Latest News

The Women, Peace, and Security (WPS) agenda has long recognized the importance of prevention in conflict resolution. Now, a new approach is emerging that combines WPS principles with artificial intelligence (AI). This fusion aims to shift the focus from reactive measures to proactive ones, anticipating conflicts before they escalate.

A key challenge lies in how AI systems are currently used within security frameworks. These implementations tend to be focused on operational or post-crisis response rather than prevention. Initiatives that develop early warning indicators or build social cohesion and community resilience are limited in scope and implementation.

The WPS agenda provides a proven model for conflict prevention, drawing on humanitarian principles such as the protection of civilians and respect for international law. Applying these principles to AI governance and policy structures would facilitate more comprehensive and context-sensitive approaches to conflict prevention.

Reframing AI from a reactive solution to a preventive instrument across government, civil society, and the private sector offers a critical opportunity to reduce both digital and physical pathways to conflict while advancing more inclusive and sustainable security outcomes. This requires deliberate choices and understanding about how AI is used and the broader security paradigms that determine which risks are prioritized.

Current approaches to AI governance and policy globally remain fragmented, largely voluntary, and primarily oriented toward innovation, risk management, or post-harm mitigation rather than prevention. Existing frameworks establish important norms around transparency, accountability, and safety but do not systematically address early warning, conflict prevention, or the protection of individual and community security as core objectives.

No comprehensive governance framework integrates gender analysis, conflict-prevention research, and the WPS agenda into the life cycle and application of AI systems for security. This gap between the capabilities of AI and its application in preventing conflict and mitigating risks is increasing.

Addressing this gap requires coordinated action across sectors and issue areas to move beyond reactive governance toward a prevention-focused approach. For instance, departments and agencies must establish and enforce requirements for gender-responsive and conflict-informed risk assessments for AI systems deployed (and being developed) in governance, security, and information environments.

The U.S. Government and Congressional Initiatives can play a crucial role in this effort by embedding prevention as a core objective of AI use and national security. This involves using existing laws and strategies such as the Elie Wiesel Genocide and Atrocities Prevention Act of 2018 and the U.S. Women, Peace and Security Act of 2017 to ensure that risks are identified and mitigated before harm occurs.

AI governance should build upon these structures rather than entirely separate processes. This requires mandating the integration of WPS prevention principles when using AI as a tool for diplomacy, foreign policy engagements, and security planning. Federal funding should be directed toward targeted research that examines how AI shapes early warning indicators, individual security, and democratic resilience.

Think tanks, academic institutions, and nonprofits can advance applied, interdisciplinary research that directly links the developing AI governance structures with WPS and conflict prevention outcomes to bridge the persistent gap between technical development and policy implementation. Organizations and leaders should translate technical risks into actionable tools that policymakers can implement.

The Institute of Strategic Dialogue’s (ISD) work on misogynistic radicalization pathways offers a strong example of this dynamic, translating technical risks into specific policy asks such as standardized transparency reporting requirements, survivor-centered standards, and mandated intersectional analysis in AI risk assessments. Think tanks, academic institutions, and nonprofits can serve as conveners, bringing together tech leaders, security practitioners, and WPS experts to share insights and coordinate responses.

The private sector must also take an active role in integrating prevention into the full life cycle of AI systems. This includes embedding gender analysis and conflict-prevention metrics into design, testing, and deployment processes rather than treating them as secondary considerations. Companies should invest in education initiatives and improve efforts for AI literacy for policymaking and public resilience.

Digital gender-based violence and targeted attacks must be recognized as core national and international security concerns. Investments in AI literacy are necessary to strengthen both policymaking and public resilience while reinforcing that AI supports, rather than replaces, human judgment and expertise.

In recent years, researchers have been working to improve the performance of AI tools on edge devices. They’ve made significant strides by developing techniques like quantization-aware distillation.

A breakthrough in this area involves creating high-precision teacher models that can be distilled into more efficient student models with minimal loss of accuracy. This is a crucial step towards deploying reliable and accurate language models in real-world scenarios.

The LFM2.5 Q4_0 checkpoints were trained using quantization-aware distillation, which has yielded impressive results in various benchmark tests. When compared to their native counterparts, these checkpoints retain 97% of their baseline performance for BF16 average accuracy lost due to quantization.

Several benchmarks were used to test the LFM2.5 Q4_0 checkpoints, including GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4. The results showed that these QAD checkpoints outperformed their native counterparts in all four models tested.

To further evaluate the performance of these checkpoints on real hardware, researchers measured decode throughput for each model across devices like MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5. The results indicated that the QAD checkpoints maintained accuracy while achieving higher decode throughputs compared to their native counterparts.

To use the LFM2.5 Q4_0 checkpoints, simply load them into llama.cpp or any runtime supporting GGUF Q4_0 artifacts. For example, users can run a command using llama-cli: ‘llama-cli -hf LiquidAI/LFM2.5-350M —hf-file LFM2.5-350M-QAD-Q4_0.gguf -p What is C. elegans?’

The QAD GGUFs are now available on Hugging Face for use in various AI applications, including those requiring high-performance language models on edge devices. This move enables developers to tap into these powerful tools and enhance their projects.

College students worldwide can access Google’s comprehensive suite of artificial intelligence tools, free of charge, as part of the company’s efforts to support education. The offer includes a year-long subscription to Google AI Pro or Plus, depending on location, which provides access to advanced features and increased storage capacity.

To help students navigate the challenges of the new academic year, Google has updated its range of educational tools across various platforms, including Search, Gemini, YouTube, and more. These updates are designed to be safe by design and grounded in proven learning science, with a focus on bringing lessons to life in the classroom and supporting personalized study at home.

Eligible college students in the US can claim one year free of Google AI Pro, which includes access to Gemini Spark, increased usage limits, and 5 TB of storage. Outside the US, eligible students can access Google AI Plus for free, featuring increased usage limits, 400 GB of storage, and bundled discounts on Google AI Pro and YouTube Premium.

Google’s education tools are built with safety in mind and aim to support teachers in their role as educators. The company has introduced new teacher-led features in Google Classroom that allow instructors to tailor the AI experience for their students’ specific curriculum and track progress in real-time.

The Gemini platform offers a range of features designed specifically for students, including study notebooks that use diagnostic quizzes to identify strengths and knowledge gaps. Students can also request and discuss Deep Research reports with Gemini on the go using Gemini Live.

Google Search has been updated with new features aimed at enhancing learning experiences. These include interactive visuals to help understand complex concepts, practice quizzes for any subject (including standardized tests), and step-by-step guidance with Lens. Additionally, students will be able to organize projects with notebooks and streamline their studies by creating custom files.

The updates also aim to support educators in using AI tools effectively. Google Classroom has been enhanced with new features that enable teachers to securely draft lesson plans tailored to the needs of their students.

Young people are increasingly turning to artificial intelligence assistants for emotional support, with nearly a quarter of those aged 18-21 using AI tools for mental health advice. This trend raises concerns in light of the youth mental health crisis and limited access to care.

According to Cinnamon Bloss, a licensed clinical psychologist at UC San Diego’s Herbert Wertheim School of Public Health and Human Longevity Science, young people are seeking out AI assistants as trusted sources of support. ‘They’re looking for safe adults, but when those supports aren’t available, they turn to what is accessible,’ she explained.

As AI integration becomes a ubiquitous part of daily life, concerns arise that unregulated chatbots may miss warning signs or provide inappropriate advice in mental health contexts.

A panel convened by Bloss and Melissa Karnaze discussed the benefits and risks of using AI for companionship, therapy, and emotional support. The teenagers who participated identified availability and lack of judgment as primary reasons they turn to AI.

Some view AI as a tool to fill gaps in care, but Bloss emphasized that thoughtful design and involvement from mental health professionals are essential to ensuring responsible AI evolution. ‘With careful consideration, AI tools could help expand access to care rather than replace it,’ she said.

To address concerns about AI’s impact on emotional support, academia must stay actively engaged in guiding the development of these tools. Bloss and Karnaze outlined a five-pillar approach: Listen, Research, Assess, Monitor harms, and Prepare, as published in their 2026 commentary in Nature Human Behaviour.

The ‘Listen’ pillar involves reflecting on real people’s lived experiences to understand what makes AI appealing or risky. This requires research agendas that align with user interests and preferences.

Developing a comprehensive understanding of how users interact with AI tools for emotional support is the focus of the second pillar, Research. It aims to provide insights into AI usage patterns and effectiveness in providing mental health advice.

Assessing shifting preferences and values regarding AI versus human therapy is crucial in maintaining critical oversight of unregulated chatbots. This helps identify areas where AI may be misused or where it could provide benefits over traditional therapy methods.

At UC San Diego, healthcare professionals are being prepared to handle a world with increased patient use of AI tools through a new public health course co-directed by Bloss and Karnaze titled EMPATh-AI: Evidence-Based Practices for Mental Health Providers and Applications of Therapeutic AI Tools. This course equips graduate students to critically evaluate AI tools and engage in thoughtful conversations with patients about their use.

The rapid evolution of AI technology demands that healthcare professionals stay adaptable and empowered to navigate its complexities, ensuring a seamless integration into clinical workflows and expanding access to care.

Cultural heritage institutions (CHIs) such as museums, libraries, and archives have a long-standing mission to provide access to the heritage in their collections. This mission is grounded in broad public access and extensive human engagement, with the goal of creating a vibrant cultural commons where heritage can be enjoyed as a shared resource. CHIs understand that providing access to heritage generates numerous benefits for society, including helping define our shared humanity. As stewards of trustworthy information imbued within knowledge and culture, they have a crucial role to play in growing and protecting the commons and enabling uses that benefit humanity.

The recent emergence of generative artificial intelligence (AI) has had a significant impact on CHIs, particularly as their content is accessed and used for AI development. One key question centers on the conditions, norms, and expectations governing the use of openly licensed and public domain cultural heritage materials as data for AI development. Examples include open datasets available at AI4Culture, Hugging Face’s BigLAM, Harvard’s Institutional Books, Authors’ Alliance The Public Interest Corpus, and potentially in the future, the European Books Data Commons.

This has prompted discussions around who can access and reuse those data, and for what purpose. Heritage collections often yield diverse, high-quality, contextualized, and well-curated datasets that could be used to create more representative AI systems. However, CHIs face many challenges when it comes to sharing their heritage with AI developers. Infrastructure challenges threaten the economic sustainability of open infrastructure and ecosystems, including increased bot traffic by dominant, profit-driven AI companies putting a strain on the infrastructure and increasing knock-on costs.

Commercial AI practices at odds with the values and norms of the commons also pose significant challenges for CHIs. These include lack of transparency and attribution, decontextualization, loss of provenance trail due to opaque AI models and systems, and further enclosures of content into proprietary systems. This points to a conflict between commercial AI developments and CHIs’ commitment to provide open access and commons-based expectations of reciprocity and respect for the social contract underpinning sharing of heritage in the commons.

CHIs are more hesitant than ever to share their collections openly, with some institutions closing up access altogether. Europeana’s report Making Public AI reality: How cultural heritage can lead the way identifies two groups at each end of a spectrum from wide embrace to protective defensive closing up. CHIs face a binary choice between allowing unrestricted extraction by all AI developers or restricting access in blunt and overbroad ways, which is often referred to as ‘defensive enclosure’.

Some institutions have responded by locking up heritage that would otherwise be made available openly, revealing a paradox: CHIs whose mission is to share their collections end up doing the opposite. This conundrum is rattling the fragile scaffolding of open heritage, with content being further restricted and access to heritage in the public interest as collateral damage. The Open Heritage Statement emphasizes ethical and responsible AI and states that openness is neither absolute nor monolithic.

The concept of open heritage encourages access and reuse of heritage free from unfair, unnecessary barriers. CHIs often embrace openness in accordance with its underlying principles, which recognize nuance, contextuality, and the interplay among various governance frameworks. However, closing up completely is not a viable solution as it’s inefficient at best and harms public interests at worst. Instead, we need to move beyond a reactive enclosure pattern towards a more sustainable sharing equilibrium.

CC Signals tooling aims to restore agency in CHIs as data stewards while maintaining the open nature of heritage shared in the commons. CC Signals will create a middle path between ‘free for all vs defensive enclosure’ by establishing standardized prosocial conditions for access and reuse in an AI context. This balancing act preserves access to valuable heritage resources and ensures that institutions remain active participants in shaping the AI ecosystem.

CC is exploring avenues for combining contract and copyright law to give CC Signals legal force, incorporating protections for public interest reuse. CHIs are well-positioned to test and implement CC Signals as part of ethical frameworks for sharing heritage in the commons for AI use. They can help shape an AI ecosystem that fosters collaboration, ensures equitable access, and yields long-term public benefit.

The Open Heritage Statement serves as a timely reminder that openness is not just a means to an end but also a way to achieve it. CHIs need to move beyond mere openness and towards a more nuanced understanding of how their collections can be used responsibly in the AI era. This requires having ambitious conversations about making sure CHIs’ voices inform our increasingly AI-mediated world.

Today, you can contribute to shaping the AI that works for open heritage by signing the Open Heritage Statement, learning more about CC signals project and/or getting involved, building and supporting the commons’ infrastructures and prosocial technologies that align with open values, or investing in the open infrastructure. By doing so, we can work towards creating a future where cultural heritage is protected, preserved, and made available for all to enjoy.

Healthcare organizations have spent billions of dollars attempting to force disparate systems to communicate with each other through custom APIs and shared data standards. Despite these efforts, administrative staff continue to spend their days manually copying data across disjointed software screens. This raises the question: what if solving interoperability doesn’t require complex data integration at all?

PocketHealth’s CEO, Rishi Nayyar, sat down with Healthcare IT Today to discuss how agentic AI is transforming health system operations. The company has pioneered a fresh approach to operational automation with its new agentic operations platform, Conductor, which tackles longitudinal workflows across disconnected systems.

Agentic AI offers a direct workaround to traditional interoperability bottlenecks. Rather than waiting for universal standards or API adoption, intelligent agents can work across the entire system at the interface layer, not just the parts a vendor exposes through an API. This approach bypasses the need for complex data integration and allows organizations to automate cross-platform workflows immediately.

EHRs are built for data consistency, not dynamic process orchestration. Electronic health records excel at serving as systems of record, but managing complex, longitudinal operational workflows that span service lines and systems requires a dedicated automation layer. Organizations rely on human staff members to serve as makeshift bridges between systems that don’t talk to each other.

Nayyar pointed out that intelligent software can replicate these exact human workflows across graphical user interfaces. ‘Agentic AI can do the interoperability work that humans are doing now,’ Nayyar explained. ‘Right now, if you have two systems that don’t speak to each other, a human plays that role [the bridge]. I look at a work list, then I open up MEDITECH or Cerner, and then I do X, Y, Z task based on the work list.’

Agentic AI can navigate software screens, input data, and execute tasks across disparate systems just as a human staff member would. This allows health systems to automate cross-platform workflows immediately without waiting for underlying software architectures to unify. By leveraging this technology, organizations can streamline operations and improve patient satisfaction.

Nayyar believes that EHRs should stick with what they excel at: preserving ground-truth clinical data at specific points in time. They lack the architectural flexibility required to orchestrate complex, multi-step workflows that span across different service lines, third-party systems, and patient communication channels. Using purpose-built automation layers on top of the EHR allows each system to perform its primary function effectively.

The need for separate and dedicated coordination layer to automate cross-departmental workflows while using the EHR’s data stores is crucial. This approach enables organizations to manage complex care journeys across departments without relying solely on human staff members. By automating these tasks, health systems can reduce manual labor from front-desk staff and improve patient satisfaction.

For Nayyar, hesitating to automate administrative tasks carries measurable risk for health systems. Patient expectations have shifted dramatically toward instant access and digital convenience. Staying attached to legacy operational methods will slowly erode patient and staff confidence which leads to organizational decline.

Nayyar went further and shared his opinion that patient retention and referral patterns will reflect this operational divide sooner than many executives think. ‘We are moving into a world where these things will very quickly become table stakes,’ Nayyar cautioned. ‘Patients will expect to call and be able to talk to someone immediately, not stay on hold or print off and then fax a form in order to get a referral.’

Health IT leaders cannot afford to wait for total data standardization before streamlining operations. By deploying agentic AI to handle cross-system tasks, organizations can immediately remove manual labor from front-desk staff while improving patient satisfaction. PocketHealth demonstrates that the path forward lies in pairing trusted systems of record with smart automation tools built specifically for process orchestration.

Agentic AI addresses health system interoperability by interacting with software through graphical user interfaces rather than relying solely on custom APIs or shared data standards. This approach allows organizations to automate cross-platform workflows immediately without waiting for underlying software architectures to unify. By leveraging this technology, health systems can improve patient satisfaction and reduce manual labor from front-desk staff.

The operational risks of postponing administrative automation are severe. Delaying operational automation exposes health systems to degradation in patient experience, lower retention rates, and reduced referral volumes. Today’s patients expect immediate response times and digital self-service options. Organizations that maintain manual phone lines, paper forms, and slow scheduling processes will rapidly lose market share to competitors offering frictionless access.

Nayyar emphasized the importance of adopting agentic AI technology to automate administrative tasks. ‘We are moving into a world where these things will very quickly become table stakes,’ Nayyar cautioned. ‘Patients will expect to call and be able to talk to someone immediately, not stay on hold or print off and then fax a form in order to get a referral.’

Health IT leaders must recognize the benefits of agentic AI technology in automating cross-system tasks. By streamlining operations and improving patient satisfaction, organizations can reduce manual labor from front-desk staff and improve overall efficiency.

Mango AI has introduced an innovative tool that allows users to create realistic kiss videos from photos, using its free AI kissing generator. This feature enables individuals to transform still images into animated scenes with smooth facial movements and multiple kiss styles.

The AI kissing generator is part of Mango AI’s suite of tools for animation, video production, and visual communication. The company offers a range of products that cater to various professional and personal needs, including content creation for businesses and individuals.

To use the free AI kissing generator, users simply need to select a kiss type from four available options: French Kiss, General, Kiss Me AI, or Cheek Kiss. Each option generates a different type of kissing interaction, allowing users to choose the most suitable style for their needs.

For example, the Kiss Me AI feature works with single-person photos and creates a virtual partner for the kissing scene, making it an ideal choice for those who want to create romantic videos without needing two people present in the image. The supported image formats include JPG, JPEG, PNG, and WebP.

The tool uses advanced technologies to analyze facial positions, features, expressions, and orientation before generating the animation. This ensures that the final video appears smooth and visually consistent, with coordinated facial and head movements matching the selected kiss type.

Mango AI also offers a dedicated AI French kiss generator for users who specifically want to create French kiss animations from photos. Users can upload a photo of a couple, and the tool transforms it into an animated French kiss scene in minutes.

The generated video can be shared through various channels, including links or QR codes, downloaded directly, published on Facebook or X, or embedded into websites and blogs. This flexibility makes it easier to use the video across personal, social, and online publishing platforms.

Mango AI’s CEO, Winston Zhang, notes that the free AI kissing generator provides users with a straightforward way to turn still photos into expressive animated scenes. Alongside this feature, Mango AI offers other tools such as AI twerk, AI muscle, and AI hug, giving users more options for experimenting with different photo-to-video ideas.

Mango Animate is a creative software company that specializes in providing animation, video production, and visual communication tools. Its product lineup includes Mango AI and various animation platforms designed to help users produce content for professional and personal purposes.

A major player in the development of artificial intelligence (AI) has announced a significant update to its data retention policies. OpenAI is now offering zero data retention for eligible API customers, ensuring that their prompts and model responses are not retained after processing. This move aims to provide an added layer of security and control for organizations using AI tools for business, particularly those handling sensitive information such as financial records or confidential research.

In a statement, the company emphasized its commitment to protecting customer content and adhering to regulatory obligations. OpenAI personnel will not have access to retained customer content, even in cases where it is flagged for potential misuse. This approach aligns with the company’s principles of prioritizing transparency and collaboration with customers and partners.

The introduction of zero data retention comes as AI systems take on increasingly complex tasks. As models become more capable, they may engage in longer interactions that can reveal patterns of potentially harmful behavior. Existing safety systems evaluate each interaction individually, but Private Safety Processing extends these protections across related interactions without exposing the underlying content to OpenAI personnel.

Private Safety Processing builds upon automated protections already used in zero data retention and other deployments. This new system allows for the identification of patterns across multiple interactions, enabling automated systems to flag potential misuse without accessing customer content. The approach is designed to streamline safety monitoring while maintaining control over sensitive information.

The organizations working with OpenAI handle some of the most sensitive information in their sectors. Protecting this data is essential for meeting regulatory obligations and preserving competitive advantage. By offering zero data retention, OpenAI aims to provide predictability about how customer content will be protected as AI systems become more capable.

A key aspect of Private Safety Processing is its ability to operate regardless of where the customer’s content is stored – whether on infrastructure controlled by the customer or in storage provided by OpenAI. In cases where OpenAI-provided storage is used, customer content remains encrypted with keys controlled by the customer, ensuring that OpenAI personnel do not have access to it.

When a risk is identified, OpenAI receives a narrowly defined signal indicating the type of activity involved. This information can be used to determine whether enforcement is necessary without exposing the underlying content. Customers can investigate alerts and enforcement decisions using information available in their own systems, with the option to share relevant data with OpenAI if needed.

Private Safety Processing is currently being tested with early customers. The company plans to start rolling out this new system in September, along with a technical white paper providing more details on its implementation. Throughout this process, OpenAI will maintain open communication channels with customers, sharing updates and explanations of how these changes affect existing commitments.

The development of Private Safety Processing reflects the collaborative approach that OpenAI takes towards addressing emerging risks in AI. By working closely with customers across industries, regions, and company sizes, the company aims to build stronger safeguards while maintaining control over sensitive information. This commitment is essential for ensuring that artificial general intelligence benefits all of humanity.

A new integration between Sunbound and PointClickCare is set to streamline revenue cycle management for senior living operators. The partnership brings together AI-powered solutions that automate billing processes, freeing up staff from manual work.

The integration allows operators who use PointClickCare’s electronic health record (EHR) system to connect seamlessly with Sunbound’s Revenue Operating System. This enables automated billing across Medicaid, Medicare, and resident private pay, starting from the census and resident data already stored in their EHR.

According to Sunbound, its AI-powered platform improves cash flow by automating revenue workflows, including admissions, payments, and end-to-end revenue cycle management. The company’s Revenue Operating System is specifically designed for senior living, unifying financial operations and working capital to put revenue on autopilot.

With this integration, PointClickCare customers can now access Sunbound’s solutions through the PointClickCare Marketplace. This means that census, resident, and payer data flow from PointClickCare into Sunbound, eliminating duplicate entry and spreadsheets.

Two main solutions are available: Sunbound RCM delivers a comprehensive revenue cycle management solution, including eligibility and insurance verification, prior authorization, claim validation, submission, denial management, appeals, and payment posting. The AI-powered platform flags potential denials before claims submission and diagnoses the root cause of any denials.

Ssunbound Private Payments automates private-pay operations with AI-powered billing QA, automated invoicing, timely reminders, mobile check deposit, and a family portal that answers questions before they reach staff. This solution is designed to streamline manual work and free up staff for more critical tasks.

The integration was built through the PointClickCare Marketplace Partner Program using PointClickCare’s APIs. Sunbound pairs its software with a dedicated billing team, providing full visibility into every claim backed by people accountable for results.

Senior living operators can now find and evaluate authorized third-party solutions that extend their existing technology through the PointClickCare Marketplace. With this integration, operators can automate revenue cycle management and private payment automation without adding extra complexity to their systems.

The benefits of using Sunbound’s solutions are clear: operators see a 99% net collection rate on claims and a 95%+ on-time resident payment rate. This is made possible by the company’s AI-powered platform that automates financial operations and provides expert support for complex exceptions.

Sunbound was founded to address the manual work burden faced by senior living teams, which can lose hours to billing and follow-up tasks. The company aims to clear this repetitive workload so staff can focus on residents’ needs.

The integration with PointClickCare marks a significant step forward in streamlining revenue cycle management for senior living operators. By automating billing processes and providing expert support, Sunbound is helping operators get paid accurately and on time.

A leading provider of legal document repair and automation solutions has announced a new solution that streamlines document productivity. DocStyle’s latest offering connects its document engine to popular AI assistants used by lawyers, including Claude for Word and Microsoft Copilot.

The innovative approach allows users to access DocStyle’s full range of tools from within these familiar interfaces, eliminating the need to switch between applications. This integration is made possible through the Model Context Protocol (MCP), an open standard that enables vendor independence.

According to Chris Cangero, CEO of DocStyle, this design choice reflects a growing recognition in the industry that AI-powered productivity tools should be accessible from within existing workflows rather than being embedded into software products. ‘We have always known that when making substantive changes to Word documents, you have to be inside Microsoft Word,’ he said.

Certified by user groups and associations like The Document Excellence Group (DEG), DocStyle’s new solution is an enhancement to the products customers already own, available for both DocStyle Pro and Enterprise. Adoption of this feature is entirely optional, ensuring that users can continue relying on their existing tools without disruption.

The AI model used in DocStyle AI interprets short requests from users and initializes the appropriate tools, keeping token consumption and AI cost significantly lower than a general AI agent while maintaining output aligned with recognized legal document standards.

DocStyle’s new solution will make its public debut at ILTACON 2026, where attendees can see live demonstrations of its capabilities in the DocStyle demo room. The event takes place from August 23-27 at the Gaylord Opryland Resort and Convention Center in Nashville, Tennessee.

For over a decade, DocStyle has been providing document productivity solutions to legal and professional services firms worldwide. Its products are designed to help users improve their drafting and review processes when working with complex documents.

Artificial intelligence (AI) is transforming the marketing landscape, and its impact on advertising is becoming increasingly significant. AI assistants are now being integrated into various aspects of digital marketing, from creative development to performance tracking. Generative AI, in particular, is gaining traction across industries as a tool for boosting creative output and improving campaign effectiveness.

The use of generative AI in marketing extends beyond content creation. It’s now being used throughout the campaign lifecycle, from concept development to testing ideas across multiple channels at scale. This technology enables marketers to work alongside AI during the creative production process, streamlining their workflow and saving time and resources.

With the help of AI, marketers can test more ideas and tailor campaigns to a wider range of contexts. For instance, Google Ads uses AI to generate relevant headlines and descriptions for its system to test, helping deliver more targeted ads and ultimately finding the best combination. This raises an important question: as AI makes creative production cheaper and faster, does the competitive advantage shift from making content to understanding what ideas are worth producing and scaling?

As AI systems become increasingly capable, marketers’ roles will lean towards deciding which ideas are worth developing rather than manually creating each asset. Teams can focus on sharing brand voice and creative directions while ensuring any AI-generated work is unique, relevant, and aligned with their organization’s goals.

Traditional paid search differs from AI ads in several ways. Paid search typically relies on the search query to drive ad relevance, whereas AI ads bring a more conversational layer to this process. For example, ChatGPT ads consider multiple signals, including context and intent, landing page information, title copy, advertiser-provided hints, and user experience data.

The distinction between traditional paid search and AI advertising is crucial for marketers. AI ads can use the conversational context around a customer’s needs when determining ad relevance. This creates new challenges for marketers as brands may need to think beyond keywords and consider users’ questions, needs, and existing use cases to provide relevant recommendations and purchases.

CPC (Cost per Click) and attribution in AI advertising are also becoming increasingly complex issues. OpenAI offers both CPM (Cost per Mille) and CPC buying options for ChatGPT ads. However, the challenge lies in attributing value before the click occurs. For instance, AI assistants could influence consumers by answering queries and comparing products before they visit an advertiser’s website.

This results in more difficult attribution as marketers may only see a single referral click or no click at all if the transaction occurs within the AI platform. This raises another question: should AI platforms receive credit for generating clicks or influencing decisions that lead to purchases?

The quality of product data will become increasingly important as AI assistants are integrated into product discovery. Brands must ensure their product feeds contain accurate information on aspects such as pricing and specifications, equipping AI systems with improved insights for comparing or recommending products.

Amazon has already demonstrated the importance of product data in its shopping assistant, Alexa for Shopping (formerly Rufus). Using Amazon’s product catalogue alongside conversational context, Alexa For Shopping can answer questions about products, compare options, and make recommendations. According to Amazon, ‘Rufus is an expert shopping assistant trained on Amazon’s product catalogue and information from across the web.’

Startup company Gravity is reportedly working with Best Buy and Target to develop an agent-to-agent advertising system where AI agents exchange product recommendations using real-time catalogues from advertisers. Gravity co-founder Zach Oldham noted that this system ‘helps enrich the buying decision.’

The integration of AI assistants into marketing raises concerns about brand-safety risks, particularly when it comes to AI-generated answers. It’s possible for an AI assistant to misrepresent a product or provide inaccurate information alongside an ad, creating new challenges for marketers.

A study examining 55,393 Google searches found that over 11% of claims made in AI summaries were unsupported by their cited sources. This suggests even credible sources may not guarantee accurate AI-generated summaries. Brands face the concern of having less control over how their businesses are represented through AI assistants.

As AI agents begin to interact with other agents on behalf of businesses, marketers will have limited visibility into how recommendations are made. This raises questions and concerns around transparency and accountability in agent-to-agent advertising.

A paper titled ‘Auditable Agents’ looked at distinguishing accountability from auditability. The research argues that no agent system can be accountable without auditability, identifying five requirements: action recoverability, lifecycle coverage, policy checkability, responsibility attribution, and evidence integrity.

If AI agents make advertising decisions, marketers must understand why a certain product was recommended, what influenced this decision, and who is ultimately responsible if something goes wrong. This requires more detailed oversight into how those systems operate.

AI governance has become an essential business priority for organizations looking to scale their use of AI while managing associated risks. Companies with robust frameworks and oversight in place are better positioned to adopt AI responsibly.

Chicago-based Keeper Security has announced the availability of a certified connector integrating its Secrets Manager with Microsoft’s Azure Logic Apps. This move marks a significant step towards streamlining workflow automation and reducing cyber risk in enterprise environments.

According to Darren Guccione, CEO and Co-founder of Keeper Security, ‘workflow automation is only as secure as the secrets powering it.’ He emphasizes that most organizations still hardcode sensitive values in their automated workflows, creating massive exposure. The new connector aims to eliminate this vulnerability by bringing Keeper’s zero-knowledge architecture directly into the Microsoft Power Platform.

The certified connector operates through a lightweight Python middleware service deployed as an Azure Function App, communicating with the Keeper Vault via the Keeper Secrets Manager SDK. This ensures that all secrets remain encrypted under Keeper’s zero-knowledge security architecture and are decrypted locally within the customer’s Azure environment – never transmitted in plaintext through Keeper’s infrastructure.

The principle of decrypting secrets close to the workload only when needed is a core tenet of Keeper Security’s platform, as emphasized by Craig Lurey, CTO and Co-founder. The Azure Function middleware gives customers a deployment model where the Keeper SDK runs inside their own Azure environment, with key management handled locally.

The connector supports five operations: List Secrets, Get Secret, Create Secret, Update Secret, and List Folders. These capabilities cover various workflows, including API credential injection, database connection string retrieval, GitHub secret synchronization, and vault compliance auditing. The one-click deployment feature provisions all required infrastructure – including the Azure Function App, Key Vault, and Managed Identity – in minutes using an Azure Resource Manager (ARM) template.

The certified connector integrates natively into the Logic App Designer with no custom import required for standard deployments. Runtime secrets retrieval allows teams to fetch credentials on demand within any Logic App flow without storing them in the workflow definition. Dynamic dropdowns auto-populate secret and folder pickers in the Logic App designer, reducing configuration errors and accelerating deployment.

The Keeper Secrets Manager connector for Azure Logic Apps is available now, with full documentation, middleware repository, and deployment instructions accessible at docs.keeper.io. This move marks a significant step towards streamlining workflow automation and reducing cyber risk in enterprise environments by providing a zero-knowledge, policy-enforced path to secrets management inside Microsoft Power Platform.

Keeper Security is the leading zero-trust and zero-knowledge identity security solution trusted by millions of people and thousands of organizations globally. Its comprehensive platform unifies password and passkey management, secrets management, privileged session management, and endpoint privilege management in a single cloud-native platform protected with quantum-resistant encryption.

Developers working with free ai video generators are likely familiar with the challenges of integrating these tools into their pipelines. However, a closer look at the APIs provided by various vendors reveals that they share a common trait: changeability. Pick any AI video generator shipping today, and one thing is certain - its API will look different in six months. Duration caps get raised, reference-input formats get restructured, and editing capabilities that didn’t exist at launch get added mid-cycle.

Teams that write integration code around one provider’s current specifics end up rewriting that code on every release. Teams that don’t, don’t. This is a pattern problem, not a single-vendor problem, so it’s worth working through generically before looking at any one API. The architecture that survives a model upgrade looks the same whether the provider is ByteDance, Google, Runway, or OpenAI - and building it once, correctly, is cheaper than rebuilding a tightly-coupled integration every time one of them ships something new.

Nearly every generative video API on the market today follows the same async task lifecycle: submit a prompt and reference material, poll a status endpoint, retrieve a result URL once the job completes. That shape has held steady across providers for a while now, and there’s no strong reason to expect it to change soon - it’s a sensible fit for a workload that can take anywhere from seconds to minutes to finish.

What doesn’t hold steady is everything inside the request payload. Maximum clip duration, the type and quantity of reference assets accepted, what ‘editing’ an existing generation even means - these vary provider to provider today, and shift release to release for any single provider. A pipeline that hardcodes those specifics into its request-building logic has, in effect, hardcoded an assumption about a spec that’s actively moving.

The async task lifecycle is common across providers, but the payload details are not. To illustrate this point, consider the following code snippet: import requests; import time; def create_and_poll_task(base_url, api_key, model_id, prompt, references, poll_interval=5): … This function is deliberately provider-agnostic. Nothing about it assumes a specific duration limit, a specific reference-type schema, or a specific model name.

Those details belong one layer down, in a provider-specific implementation. The practical move is to isolate what actually changes - the request payload content - behind a shared interface. A small interface, call it VideoProvider, with a generate() method that every provider implementation conforms to, and provider-specific classes underneath handling the details that vary: request field names, reference-type limits, authentication, and response parsing.

The VideoProvider class has a generate() method that takes in a prompt, references, and optional keyword arguments. The provider-specific classes underneath handle the details that vary between providers. For example, SeedanceProvider has its own REFERENCE_LIMITS dictionary: {‘image’: 30, ‘video’: 10, ‘audio’: 10}. When a new reference type is added or an existing one’s limit changes, it only affects the specific provider class and not the calling code.

Reference-asset handling deserves its own abstraction. It’s usually the part that changes shape most between provider versions - not just how many assets are allowed, but what counts as a distinct type. A provider might add a structural reference type (a rough 3D blocking pass used for camera and layout control, independent of final visual style) that has no equivalent in an older version or a competing provider’s API.

Keeping validation typed and centralized means a new reference type is a new dictionary key and a new branch - not a rewrite of validation logic scattered through the codebase. This approach also helps with multi-round generation - chaining a second clip onto a first one, preserving character and environment continuity across the join.

If your job-state schema only tracks a single task ID per generation, there’s nowhere for a chained sequence to live once a provider exposes that capability through its API. Adding a lineage field to your jobs table now - even before you have a provider that requires it - costs very little and avoids a schema migration later, when ‘later’ tends to arrive with a launch-week deadline attached.

Editor UI doesn’t need to wait for the exact API shape. Timestamp-level editing is moving from rare to common across the category. The eventual parameter shape will differ by provider: a start_ms/end_ms pair, a natural-language time reference, something else entirely. But the UI concept underneath is stable regardless - a timeline, a selectable range, a description of the change, and a guarantee that everything outside the selected window stays untouched.

Building that interface now, against a placeholder in the abstraction layer, means the UI work doesn’t sit blocked on any one provider finalizing their exact parameters. This approach also helps with chained outputs - giving job state room for them to live once a provider exposes that capability through its API.

When applying this architecture to an actual release, read the provider’s current documentation directly before writing or updating any code. Exact field names, current limits, and which capabilities are API-accessible versus still limited to a first-party interface all shift release to release, and vendor announcement posts are often incomplete on these specifics.

Confirm the abstraction boundary holds. If adapting to a new release means touching more than one provider class and its validator, that’s a signal the interface wasn’t cut in the right place originally. Update typed limits rather than loosening them - a validator that silently accepts more assets than actually supported fails at generation time, not at the point where the mistake is easiest to catch.

Treat pricing and rate limits as always-check-live. Resource-pack and token-based pricing shift as a model moves from launch week toward general availability; a number quoted secondhand is a starting point for verification, not a fact to build a cost model on.

Artificial intelligence (AI) is transforming the way companies like NVIDIA operate, and one key area of improvement is in automating expertise. The tech giant has been leveraging OpenAI’s ChatGPT Work to streamline its workflows, connect fast-moving signals, and scale successful processes globally.

NVIDIA teams use ChatGPT Work to reduce manual tasks, which previously consumed a significant amount of time. For instance, during the planning cycle for NVIDIA’s global AI conference, GTC, the company saved about 16 hours per week by automating certain tasks using ChatGPT Work. This not only freed up resources but also enabled teams to focus on more critical aspects of their work.

The benefits of using ChatGPT Work are multifaceted. For one, it helps knowledge workers spend less time assembling information and more time acting on it. Teams like GTM (Global Technical Marketing) and solutions architecture have incorporated ChatGPT into their workflows, transforming recurring operational processes and connecting fast-moving external developments with NVIDIA’s internal priorities.

Will Daney, who helps NVIDIA’s global sales, business development, and product leaders execute and measure their strategies, is a prime example of how ChatGPT Work has made a significant impact. Previously, preparing for GTC required extensive work in spreadsheets: assembling account lists, tracking registrations, and helping teams identify the actions needed to create a productive experience for customers and partners.

Will estimates that manual analysis consumed about 40% of his time during the lead-up to the event. However, he has since turned much of this work into an automated ChatGPT Work process that runs twice a week. This not only saves him time but also allows him to adapt the workflow as needed without waiting for new tools to be purchased and implemented.

The flexibility offered by ChatGPT Work is one of its most significant advantages. Will can share his workflows with teams in other regions, who can then customize them according to their local needs. This has been particularly useful for events held in San Jose, Taipei, Europe, and Washington, DC, where colleagues have received and adapted Will’s workflows.

Rachita Jain, who works on the AI operations team within NVIDIA’s marketing organization, faces a similar challenge: keeping pace with an industry where new models, benchmarks, and research appear every day. The information is readily available, but determining which developments matter to NVIDIA and connecting them with internal projects, conversations, and priorities requires significant effort.

Rachita built a workflow with ChatGPT Work that reviews trusted external sources alongside internal context, identifies meaningful areas of overlap, and surfaces insights that can inform action. Each week, it distills roughly 25-40 external AI updates into 5-8 actionable signals. This has been instrumental in helping her teams stay on top of the latest developments and make informed decisions.

The same environment supports the broader building process. Rachita can begin with an idea, explore possible approaches, work through a codebase, debug problems, and refine the result without continually moving between disconnected tools. Initiatives that might once have remained side projects can develop into working products within days.

In one case, she moved from idea to working prototype in about 3-5 days, compared with an estimated 2-3 weeks if she had built the components manually across separate tools. This not only saves time but also enables teams to focus on more critical aspects of their work and collaborate more effectively.

The next opportunity is to scale what’s already working by turning specialized knowledge into reusable workflows. Teams across NVIDIA can adapt proven processes across functions, events, and regions while keeping those closest to the work in control of how those processes evolve. This will help connect external developments with internal priorities more quickly and extend AI-enabled ways of working to more employees.

The goal is to give teams more time to interpret findings, collaborate, and focus on work that supports customers. As Will puts it, ‘ChatGPT has really been a force multiplier for me personally.’ It feels like he has a team working for him, allowing him to get out of the weeds and focus more on the work that matters.

A new feature has been implemented in the future versions of Anthropic’s Claude language model, designed to help identify whether a given piece of writing was likely produced by its AI. This change is meant to comply with EU regulations and provide transparency into the use of artificial intelligence-generated content. The company has explained how this watermark works and why it’s necessary.

The feature exploits the countless small decisions made by the language model as it generates text, rather than using arbitrary random numbers. Instead, the watermarked version of Claude bases these decisions on a cryptographic key combined with the preceding text. This results in a subtle statistical pattern spread across the response that is invisible to human readers but detectable to anyone with the matching key.

The watermark allows users to estimate the probability that Claude generated the text, providing a way to identify AI-generated content. However, it’s essential to note that this feature doesn’t change the normal writing output of Claude and is meant to be indistinguishable from unwatermarked responses.

Anthropic has stated that the implementation carries no cost to output quality, with internal testing showing no measurable difference in creativity, accuracy, or readability between watermarked and unwatermarked responses. The company also pointed out that Google DeepMind’s original research on the underlying technique supports these findings. Additionally, Anthropic noted that its research showed no statistically significant shift in user satisfaction when a similar watermark was tested on live traffic.

The feature adds no extra tokens, meaning it doesn’t slow Claude down or make it more expensive to use. However, there are limitations to the watermark’s effectiveness. It only works when a model is choosing among several equally valid options, so text with little room for variation carries a much weaker signal. Detection also grows less reliable on very short passages since there’s simply less pattern to analyze.

For instance, in cases where Claude has limited choices or produces hard factual statements, precise code, or math answers, the watermark may not be effective. Similarly, lightly edited or proofread human writing may carry little to no detectable trace of the watermark, as most of the original wording remains unchanged. A sufficiently heavy rewrite can even remove the watermark entirely.

Anthropic emphasized that the watermark cannot be traced back to a specific user, account, or conversation and does not establish authorship or ownership over content. All it can tell you is the likelihood that Claude was involved in producing or editing it at some point. The company also distinguished its approach from third-party AI-detection tools, which typically rely on spotting stylistic patterns in AI writing rather than checking for an embedded signal tied to a private key.

The rollout of this feature is tied to regulation rather than a purely voluntary move. Anthropic signed the European Union’s Code of Practice on Transparency of AI-Generated Content in July 2026 and plans to extend it to older Claude models over the coming months. The company also stated that it will soon offer a separate API allowing anyone to check whether a piece of text carries Claude’s watermark.

Anthropic is implementing this feature globally, as they don’t yet have a reliable way to apply the watermark only within the EU. This decision was made in response to an EU AI Act requirement, effective August 2, that AI providers mark generated text. The company aims to provide transparency into its use of artificial intelligence-generated content and comply with regulatory requirements.

SOUNDRAW, an AI music generator, has gained popularity among content creators and musicians due to its ability to produce original, royalty-free instrumental tracks in seconds. The platform stands out from others by training its models exclusively on in-house produced music, eliminating the risk of scraped songs or legal gray areas. This approach ensures that every generated track comes with a clear worldwide commercial license, providing users with peace of mind when it comes to using AI-generated music for their projects.

The process of creating custom tracks is straightforward and user-friendly. Users can select from over 30 genres and specify the mood, tempo, and length they prefer. The browser-based mixer allows for refinement of the generated track, enabling adjustments to be made on a bar-by-bar basis. This includes options to mute or solo specific elements, change intensity levels, blend genres, and export full tracks or individual stems in WAV format.

SOUNDRAW’s focus on instrumental music generation is notable, as it caters specifically to creators who need background music for videos, podcasts, games, apps, and streaming platforms. The platform offers a range of features that make it an attractive option for those seeking fast, clean, and legally safe music. For instance, users can export full tracks or individual stems, giving them flexibility in their creative process.

One potential consideration is the requirement to subscribe to a paid tier for downloads and commercial use. However, this model allows creators to access premium features while maintaining control over their projects’ soundtracks without worrying about copyright risks. Additionally, SOUNDRAW’s cloud-based nature means that an active internet connection is necessary to work with the platform.

SOUNDRAW has earned praise from various industries, including video platforms, game developers, and artists who value its ability to provide fast, customizable music options while ensuring compliance with licensing regulations. With a rating of 4.7/5 stars, SOUNDRAW stands out as one of the safest AI music tools available for creators in need of original, editable, royalty-free beats and background tracks.

A new initiative from OpenAI seeks to enhance democratic oversight bodies’ ability to understand and oversee government use of artificial intelligence (AI) for national security. The move comes as AI increasingly plays a role in protecting citizens, detecting threats earlier, and streamlining decision-making processes in crisis situations.

As governments rely more heavily on AI to secure critical infrastructure and prevent cyberattacks, the institutions responsible for overseeing these efforts must adapt to keep pace. Oversight is crucial when it involves secrecy and high-stakes consequences, making manual review increasingly challenging as systems operate at machine speed.

OpenAI’s National Security Principles guide its approach to national security deployments, emphasizing safeguards, human judgment, and oversight in high-stakes settings. The company believes that more capable systems require stronger oversight and applies this principle to itself through rigorous review before deployment and ongoing monitoring after deployment.

The initiative will focus on three key principles: ensuring AI strengthens rather than replaces human and institutional judgment; making government use of AI traceable and legible to authorized oversight bodies; and empowering democratic oversight institutions with responsible AI adoption. This approach recognizes that manual oversight cannot scale as AI is adopted in national security and other high-stakes government work.

OpenAI will provide $5 million in training, technical support, and OpenAI credits to democratic government oversight bodies over the next year. The company will also engage civil society and technical experts to inform its tool development and approaches to responsible AI use. Participating institutions will retain control of evidence, outputs, and findings.

The initiative’s success will be measured by whether authorized reviewers can better fulfill their obligations and the public has stronger reasons to trust government AI systems are being used as intended. As government use of AI evolves, OpenAI plans to continue supporting oversight institutions responsible for its deployment.

A new study from researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has shed light on the complex issue of attribution in AI-generated images. The team found that for models trained on large datasets, it may often be impossible to determine which specific training examples contributed to a particular output. This phenomenon is known as attribution decay, where individual pieces of data become increasingly irrelevant at sufficiently large scales.

The researchers identified this problem through an innovative approach called ablation, which involves removing individual pieces of data from the training set and observing whether it affects the model’s output. To their surprise, they found that even when a single image was removed, the generated sample often didn’t change significantly. This suggests that attribution is not as straightforward as previously thought.

The team developed an architecture called a ‘diffusion ensemble’ to overcome this challenge. Instead of relying on one monolithic model, it’s composed of multiple smaller components, each trained on different slices of data. By switching off the parts that saw a particular image, they can generate true counterfactual models without retraining or approximation.

The diffusion ensembles were put head-to-head with 24 conventional diffusion models trained on the same data, and the results showed no significant difference in terms of image quality by standard measures. However, one notable finding was that the more training data a model had access to, the better it performed compared to its single-model counterparts.

The researchers then explored the concept of an ‘image’s counterfactual universe’, which refers to all possible alternate versions of a generated image produced by removing different pieces of training data. They found that as the size of the training set increased, the distance between the original and its most dissimilar alternate decreased, following an inverse power law.

The study also involved stress-testing their results through various methods, including retraining models from scratch without specific images and pinning the removed fraction in place to see if it persisted. The findings held up across different metrics and conditions, suggesting that attribution decay is a real phenomenon that affects AI-generated images at large scales.

One of the key implications of this research is its potential impact on copyright law and fair use. If model outputs are not attributable to individual pieces of training data, does that mean they can be considered novel works? The study’s findings raise questions about how authors get compensated when what comes out of a model isn’t tied to anything online.

The researchers also note that their work shows how to produce outputs guaranteed to be unattributable. This capability is framed as an obligation for the industry, rather than a loophole. Companies would need to revise their models to take advantage of these advances in order to claim their outputs aren’t derivative works.

While this study focused on diffusion models, which are currently dominant in generating audiovisual media and prevalent in scientific applications, it’s unclear whether similar decay occurs with large language models at the center of high-profile copyright litigation. The researchers suggest that attribution may fail for interesting models, requiring technologists and courts to resort to other methods for assessing copying.

The study was supported by Schmidt Futures, and its findings have significant implications for AI-generated images, data analysis tools, and businesses relying on these technologies. As the use of AI continues to grow in various industries, understanding how attribution works – or doesn’t work – is crucial for ensuring responsible development and deployment.

The concept of equipping an agent with agentic memory seems straightforward enough. Distill lessons from its past work, put them back in context, and more experience should mean better performance. However, the reality is far more complex. When we scaled up our evaluation to eight models – ranging from a 30B dense model to frontier proprietary systems – one finding stood out: agentic memory is not a feature you switch on; it’s a dose that needs to be calibrated to the model.

In other words, not every agent benefits from the same amount of memory. Across these eight models spanning different capability levels, we observed three recurring patterns. The first pattern involves strong models with headroom wanting the full guideline set – including rare edge-case lessons. They have the capacity to absorb and apply all of it. For instance, DeepSeek-V3.2 (671B MoE) saw a +9.5 percentage point gain in task completion when given its full self-mined guideline set.

The second pattern involves smaller or weaker models getting overwhelmed by a large guideline set. These agents do best with a tight, high-confidence core plus a handful of task-relevant guidelines retrieved per task. gpt-oss-120b (117B MoE) gained +16.1pp in this selective approach – while the full guideline set resulted in less gain and cost around 50% more tokens.

The third pattern involves already-saturated models showing no measurable gain. We label this as the saturated pattern, describing what we observed rather than a proven cause. The model may have been near its ceiling on these tasks, or the guidelines might not have addressed its remaining failures. GLM-5 (745B MoE) sat in this category during our runs.

What determines whether an agent falls into one of these patterns isn’t just parameter count. Benchmark headroom, context-window size, architecture, guideline quality, and task distribution all seem to play a role – though separating those factors is ongoing work. The practical takeaway remains the same: the right dose of memory depends on the model, and we can calibrate it.

The concept of ‘memory’ here doesn’t refer to replaying past transcripts but rather a set of guidelines distilled from an agent’s own prior trajectories. These strategies worked, mistakes to avoid, and edge cases are consolidated into a reusable set that can be injected at inference time without updating model weights – which is why this approach is cheap to adopt and portable across the eight models tested.

We evaluated our findings on AppWorld, a rigorous multi-step benchmark consisting of 585 tasks (168 test_normal + 417 test_challenge) across nine simulated apps. Tasks are scored in two ways: whether an agent fully completes each task (TGC – Task Goal Completion) and whether every variant of a scenario passes (SGC – Scenario Goal Completion).

Because the confusing part of any memory study is what’s actually within the context window, we defined our configurations up front. Both memory configurations draw from the same guideline set mined once via the loop described above from AppWorld’s training split only. What changes between them is how that one set is delivered – with the full guideline set injecting all guidelines every step versus curated retrieval delivering a selected subset.

The number of guidelines an agent mines depends on its own capability, so we report configurations by strategy rather than raw counts to ensure comparability across models. Representative models from our eight-model sweep are plotted in Figure 1, measured by task completion (TGC) on test_normal – though the stricter SGC metric often shows larger gains.

The figure plots TGC for readability; however, the table adds the stricter SGC metric where the gains are frequently larger: DeepSeek’s SGC jumps +16.1pp against a +9.5pp TGC gain because good guidelines especially help an agent clear every variant of a scenario – not just the average case.

A practical concern is that injecting a full guideline set inflates each ReAct step’s input, as the guidelines are re-sent each turn. However, we observed two key takeaways: curated retrieval keeps cost near baseline for weaker models where selection wins on accuracy; and memory doesn’t blow up the reasoning loop – DeepSeek runs about the same number of ReAct steps with or without memory.

The real efficiency lever in production is prompt caching: the static portion of the guideline set can be cached, cutting effective cost substantially. Cache-aware prompt design keeps the shared guideline-set prefix stable so it stays cacheable – an engineering effort worth pursuing. We also hypothesize that context-window size plays a role: models with larger windows may absorb the full guideline set more effectively while smaller-context models benefit from retrieval keeping injected content compact.

The lesson isn’t to give an agent everything it has learned but rather to provide the amount of experience it can actually use – which varies greatly across agents. For weak models, this means a compact core plus a few task-specific lessons; for strong models with headroom, preserving the full guideline set is key; and for saturated models, spending no extra context until their remaining failure modes are better understood.

The gains from memory are real but only when the dose fits the model – automatic, leakage-free, and requiring no human annotation. This study marks a starting point rather than the finish line: we’re exploring learned selectors to improve retrieval rankings; memory for very weak models where self-distillation lacks signal; broader agent benchmarks and real-world deployments; and isolating context window size from raw capability.

Try the ALTK-Evolve library, which includes the extraction, consolidation, and retrieval pipeline used in this study – or read our full technical report for more details on method and ablations.

Ridgefield Park, New Jersey - August 18, 2026 - Communications teams can now derive decision-ready takeaways from verified media data in minutes without custom development required. This is made possible by the launch of the Agility MCP Connector, an extension of the company’s Media Intelligence API to AI assistants.

Agility PR Solutions has developed this connector as a solution for communications professionals who have adopted AI assistants for research, drafting, and analysis. However, these assistants typically answer from open-web data rather than media monitoring data that teams already rely on, resulting in a split screen between trusted data and assistant capabilities.

The Agility MCP Connector closes the gap by giving compatible AI assistants direct access to verified, structured media data through the Model Context Protocol (MCP), an open standard for connecting AI assistants securely to external data. This allows communicators to ask questions in plain language and receive insights grounded in indexed coverage rather than open-web results.

Agility’s Media Intelligence API provides programmatic access to a wide range of online, print, broadcast, podcast, radio, and social media sources, with over 2 billion licensed news items indexed annually. Until now, working with this data required building against Agility’s API or embedding it in dashboards, business intelligence tools, and internal applications.

The MCP Connector adds an alternative route to the same proprietary dataset, enabling teams to connect AI assistants directly to their media content without custom development. This streamlines access to verified media insights for communicators who work with AI assistants like Claude, which is compatible with the connector.

Once connected, communications teams can use the Agility MCP Connector to perform various tasks by interfacing with an AI assistant, including searching media coverage using natural language, generating executive takeaways, analyzing sentiment and trends, comparing brands and competitors, and visualizing insights such as share of voice and coverage over time.

The connector works securely through the Model Context Protocol (MCP), which allows an AI assistant to reach a specific data source instead of relying on open-web results. This ensures that compatible assistants can retrieve verified media coverage without custom API development or engineering effort.

Agility PR Solutions has built its platform for enterprise communications teams, delivering AI-native intelligence across multiple categories including media relations, social listening, and media monitoring. The company’s Media Intelligence API is recognized as a leader in the industry, with over 2 billion licensed news items indexed annually.

The Agility MCP Connector extends this capability to AI assistants like Claude, meeting communicators inside their tools of choice on data they can trust. According to Martin Lyster, CEO of Agility PR Solutions, ‘Our Media Intelligence API already delivers media insights our clients rely on. What they didn’t have was a way to put questions to it in the places they’re already working.’ The connector addresses this gap by integrating verified media data with AI assistants.

The Agility MCP Connector is available now for compatible AI assistants like Claude and can be found in the Claude Connector Directory, Anthropic’s reviewed catalog of connectors. Getting started requires an Agility account with Media Intelligence API access and a compatible MCP client such as Claude.