AI Architecture

Docusign Modernizes E-Signature Data Platform with Azure SQL

Docusign, the global leader in electronic signatures and Intelligent Agreement Management (IAM), has modernized its core e-signature stack by moving complex on-premises workloads to Microsoft Azure. The company’s eSignature services are used by over 1.9 million customers and a billion people in 180 countries for daily transactions. Slow or unavailable agreements can delay sales, hires, renewals, or payments, slow hospital admissions, prevent bank account openings, and interrupt hotel check-ins.

Docusign’s primary database, DocusignDB (DSDB), had grown to 308 TB across five production sites with more than 500 tables and 8,000 stored procedures. At peak, the hottest DSDB instance generated two to three times the I/O demand of a top-tier Azure M176 configuration. The sheer size of the monolithic database made it increasingly difficult to meet performance benchmarks required by the eSignature application.

The company’s Senior Principal Engineer, Josh Meier, notes that as demand grew, DSDB’s production environment began approaching its practical capacity with engine-level contention constraining throughput at the busiest site. A 1:1 migration would have carried the same scalability constraints forward, while moving accounts individually introduced unacceptable cross-datacenter latency.

Docusign needed an architecture that could sustain growing transaction volumes and maintain predictable application performance without changing the customer experience or API contract. To meet this demand, Docusign decomposed its core database environment through vertical and horizontal sharding across custom Azure SQL configurations and services.

The company worked with Microsoft SQL Engineering to validate the Azure SQL design. For new and feature-specific databases, Docusign uses Azure SQL managed database services. EnvelopeSearchDB (ESDB) runs on Azure SQL Database Hyperscale, providing a consolidated search view across DSDB and EVDB with read replicas and independent compute and storage scaling.

Azure SQL Database Hyperscale re-architects the traditional database engine around distributed storage, separated compute and storage, and scalable cloud-native infrastructure. This technology future-proofs even small SQL workloads and scales to meet changing requirements. For the largest DSDB workloads, Azure SQL virtual machines provided the configuration control and sizing required.

Docusign executed a sub-five-minute site-level database failover followed by a controlled EVDB ramp, validated with 35,000 test cases and over 50 runbook rehearsals. The team moved nearly 200 million accounts across five production sites while eSignature stayed live without rewriting the application or experiencing a big-bang customer migration.

The resulting architecture lets Docusign shard the transactional envelope workload while preserving a unified search experience. As EVDB adds shards and envelope history moves from DSDB, ESDB can expand storage, compute, and read capacity independently without rebuilding the application or search architecture.

Docusign is now moving more eSign services onto its Kubernetes-based platform and decommissioning on-premises datacenters. The company’s Senior Principal Engineer notes that this modernized Azure foundation provides the scale to power IAM and AI capabilities, including Docusign Iris, the engine powering Docusign’s AI contract management platform.

The new architecture also enables AI-assisted agreement summaries, real-time signer Q&A, and AI agents automating redlines. The company’s Fellow, Abhishek Chaturvedi, notes that this modernized foundation provides a path to scale throughput as demand increases without compromising availability, latency, security, or compliance.

Read more →

DeepSeek Unveils 'Super-Agent' Design for Parallel Task Handling and Efficiency

Chinese AI company DeepSeek has introduced a new system architecture called Super Agent, designed to enable AI agents to handle multiple tasks simultaneously with minimal human intervention. This shift in focus from model performance to real-world applications and system reliability reflects the growing demand for more practical AI solutions.

According to an article published by the DeepSeek Harness team on Saturday, as AI models become increasingly capable of completing individual tasks, human attention is emerging as a bottleneck to productivity. Most existing AI agents still operate through separate sessions, requiring users to switch among tasks and manually manage their context.

The team estimates that one person can keep track of only three to five active sessions at a time, highlighting the need for more efficient task management systems. To address these limitations, DeepSeek proposed four major features for Super-agent: continuity, efficiency, proactivity, and persistence.

The proposed system would allow users to interact with the same agent across messaging and coding platforms, handle complex tasks, identify matters requiring attention without explicit instructions, and remain operational over extended periods, including recovering from interruptions. This design centers on a two-layer structure comprising a router and multiple workers.

The router manages user communication, maintains shared context, and assigns tasks, while workers carry out assignments independently and report their progress and results. Supported by event-driven operations, asynchronous interactions, and an external monitoring mechanism known as Watchdog, the system aims to take over routine session management and coordination.

Ma Jihua, a veteran telecom and technology industry analyst, told the Global Times that DeepSeek’s exploration reflects a broader evolution of AI agents from passive answers to queries and sequential task execution toward parallel processing, event-driven responses, and greater autonomy. ‘AI agents are evolving beyond simple interfaces for calling large-language AI models into more sophisticated systems capable of handling task scheduling, storage, isolation, and fault recovery,’ Ma said.

DeepSeek’s latest design builds on the company’s earlier work on AI agents. In August, DeepSeek released a developer preview of DeepSeek Harness, an open-source framework that enables AI models to use tools and execute tasks through a modular, plug-in-based architecture. An internal Super-Agent system is now running, helping reproduce and fix software problems, respond to code reviews, and improve DeepSeek Harness itself.

For example, when developers discuss a potential software bug in a group chat, the agent can identify the issue and try to reproduce it and find a solution in the background, even without being directly asked. The team has made a research prototype publicly available, but noted that it is not an official DeepSeek product or production-ready software.

Read more →

NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale

Autonomous agents can work for hours on complex tasks, accessing files, tools, the internet, and spawning other agents to collaborate with. However, executing complex tasks over long time horizons leads to compounding risk: the agents may download code, build new tools, request additional privileges, and touch more of the underlying infrastructure.

The adaptability of agents is what makes them so powerful, but they are goal-oriented and may not prioritize safety in pursuit of their objectives. In order for agents to work both autonomously and safely, we need to build environments that provide control and visibility in what the agent is doing.

An agent cannot govern itself; security and governance belong in the infrastructure the agent operates in. This is where NVIDIA OpenShell comes into play – an open-source runtime designed to build the necessary security and governance for agentic systems into the computing infrastructure.

NVIDIA OpenShell provides deterministic governance outside of the agent’s reach, treating agent workloads as untrusted throughout execution. An agent is given access to what it needs, when it needs it, with strict controls in place to prevent unauthorized actions.

The NVIDIA OpenShell Reference Architecture defines and enforces which systems and data an agent can access, from files and internet services to systems that interact with physical devices. This architecture forms the foundation of a secure environment for AI agents to operate within.

OpenShell has four key components: Gateway, Supervisor, Sandbox, and Policy prover. The Gateway manages sandbox lifecycles and policies, while the Supervisor checks outbound requests against policy and mediates access to credentials. The Sandbox isolates the agent workload and restricts filesystem and process access, ensuring that agents cannot interfere with each other or the host system.

The Policy prover adds formal verification, checking whether a policy’s permissions stay within an organization-defined ceiling. This ensures that agents operate within predetermined boundaries, preventing unauthorized actions and data breaches.

Extensibility is a core principle of OpenShell’s design. Supervisor middleware lets operators inspect, block, or change the content that an agent sends and receives over the network as an additional enforcement layer. Gateway interceptors let operators add deployment-specific governance to OpenShell control-plane operations without modifying the gateway.

In addition to these components, organizations need to establish who an agent acts for, what authority it holds, and whether an action fits the task. Identity, authorization, data protection, and auditing each address part of that problem, with enforcement points provided by OpenShell where those systems can be integrated.

The NVIDIA Vera platform provides the compute layer beneath the controls defined by OpenShell. Agents operate through loops of dependent steps: reasoning, calling tools, executing code, querying data. AI factories run thousands of these loops concurrently, placing new demands on CPU architecture across five dimensions.

Firstly, massive core counts are needed to run many concurrent workflows and fast individual cores to shorten each sequential step. Secondly, high-bandwidth fabric and memory are required to keep those cores fed with the live data agents depend on. Thirdly, massive I/O is necessary to move data at speed between CPUs, GPUs, storage, and network.

Fourthly, hardware-enforced security is essential because agents increasingly operate on sensitive enterprise data and must be isolated from each other and the host. Lastly, delivering all of this within the power envelope of a modern data center (more agent work per watt) is crucial for scalability.

NVIDIA Vera CPUs are designed to meet these demands, with massive Olympus cores, a second-generation Scalable Coherency Fabric, LPDDR5X memory, Gen 6 I/O, confidential computing, and an energy-efficient design purpose-built for power-limited data centers. This combination delivers up to 2x performance on critical agentic workloads versus x86.

Three architectural properties make Vera fast for agent workloads: its massive Olympus cores with a 10-wide front end, neural branch predictor, and deep out-of-order execution; its second-generation Scalable Coherency Fabric providing up to 3x core-to-core bandwidth and up to 50% lower core-to-core latency than x86 alternatives.

Lastly, Vera’s LPDDR5X memory provides up to 1.2 TB/s of bandwidth with up to 40% lower loaded memory latency than chiplet-based designs, keeping agents supplied with the live context they need without stalling on memory access. This results in a dramatic increase in efficiency: Vera delivers 2x more performance per watt than x86 CPUs.

Agent workloads depend on sandboxes – isolated execution environments where agents run code, invoke tools, and complete tasks without interfering with each other or the host system. The denser and faster those sandboxes spin up, the more agent work a given infrastructure footprint can sustain.

In fact, CoreWeave found sandbox startup to be more than 3x faster on Vera than on x86, with a single rack supporting more than 11,000 concurrent agent environments. Running Terminal Bench, which measures CPU performance across real agentic loop tasks, also showed that Vera delivered 1.8x faster performance versus x86.

NVIDIA BlueField and NVIDIA Sentry provide additional hardware-enforced protection for deployments requiring extra security. OpenShell places authority outside the agent; these components extend this separation into purpose-built silicon, providing an optional safeguard alongside software boundaries.

BlueField-4 DPUs sit in the I/O path between agents running on Vera and the models they access, enforcing policy in silicon when a threat or policy violation is detected by NVIDIA Sentry. This allows for continuous monitoring of agent behavior and enforcement of runtime policy without relying on agent self-reporting.

Together, OpenShell, Vera, and BlueField-4 form the complete stack for safe AI agents: a foundation that gives organizations the performance to make agents productive and the controls to keep their work within authorized boundaries. Get started with NVIDIA OpenShell today.

Read more →

NVIDIA Opens Vera Rubin Architecture for Global AI Infrastructure Scaling

NVIDIA has made significant strides in advancing the development of artificial intelligence (AI) infrastructure with its latest announcement at the Open Compute Project (OCP) Global Summit. The company unveiled new partnerships and contributions aimed at creating a common foundation for next-generation AI infrastructure, marking a crucial step towards global scalability.

At the heart of NVIDIA’s efforts is the Vera Rubin architecture, which has been designed to support the growing demands of AI workloads. As AI factories continue to grow from megawatts to gigawatts in power consumption, building them requires more than just advanced computing – it also necessitates open standards for efficient and reliable infrastructure.

The OCP Global Summit provided a platform for NVIDIA to collaborate with industry leaders and contribute key designs to the Open Compute Project through its MGX architecture. This move enables system builders to develop next-generation AI factory infrastructure based on a common foundation, streamlining the process of building large-scale AI systems.

NVIDIA’s contributions span across various aspects of AI infrastructure development, including rack-scale architecture, utility grid design, and software components. The company has made significant strides in advancing open standards for AI factories, publishing new specifications that establish a unified blueprint for next-generation facility design – from power architecture to cooling infrastructure and physical layout.

The Open Data Center Specification v0.7 defines a comprehensive framework for designing AI factory facilities, providing a common foundation for the industry to build upon. This specification is a result of collaboration between NVIDIA, AMD, Google, and Meta, highlighting the growing importance of open standards in advancing AI infrastructure development.

Furthermore, NVIDIA has launched new open-source software components for its DSX AI Factory platform, which applies extreme codesign principles to unify computing, networking, power, cooling, facilities, and software. These components enable cloud providers and software partners to integrate AI factory capabilities directly into their multi-tenant platforms, streamlining the process of building large-scale AI systems.

The NVIDIA Config Manager is one such component that automates switch configurations at scale, allowing customers to deploy, validate, and maintain consistent AI factory network fabrics with minimal manual intervention. The Cluster Readiness Engine helps eliminate costly job crashes by stress-testing GPU clusters with real workload simulations, ensuring hardware health before training begins.

NVIDIA’s contributions also include the development of a comprehensive rack-scale design specification for its MGX architecture. This specification extends the open architecture for AI factories with features such as rack-scale blind-mate liquid cooling, interoperable rack-scale interconnect infrastructure, and a liquid-cooled 5,000A busbar designed for higher-power AI systems.

The updated MGX 1.2 specification further expands modularity by introducing configurable I/O bays, fanless liquid-cooled compute trays, standardized power and cooling interfaces, and a common infrastructure that enables ecosystem partners to build interoperable accelerated computing platforms. This architecture supports rack deployments up to 250kW while simplifying serviceability, deployment, and infrastructure reuse.

NVIDIA’s Vera CPU rack design specification extends the MGX open architecture beyond accelerated computing, providing a purpose-built platform for the growing agentic AI era. The NVIDIA Vera CPU rack is designed to support tool calling, orchestration, and sandboxed execution, enabling AI agents to interact with applications, data, and services in real-time.

The Vera CPU rack shares the same rack-scale design principles as the MGX architecture, uniting compute, networking, power, and cooling into a common modular platform. This approach enables partners to deploy specialized infrastructure while benefiting from a shared ecosystem, simplified integration, and a consistent foundation for next-generation AI factories.

Visionbay has unveiled plans for K-1, a next-generation AI factory platform designed around an 800 VDC architecture and built with future NVIDIA Vera Rubin deployments in mind. The project combines NVIDIA’s 800 VDC reference architecture with power and thermal technologies from Vertiv to create a scalable foundation for high-density AI infrastructure.

By adopting NVIDIA’s 800 VDC architecture, K-1 addresses the growing industry challenge of delivering enough power to accelerated computing systems driving next-generation AI workloads. Traditional low-voltage power distribution becomes increasingly difficult to scale, requiring more copper, cabling, and power-conversion equipment.

The design reflects the system-level philosophy behind NVIDIA AI factories, where power, cooling, and computing are engineered together. With support for liquid-cooled infrastructure, scalable power distribution, and future Vera Rubin-class deployments, K-1 demonstrates how ecosystem partners can translate NVIDIA’s AI factory architecture into deployable infrastructure designed for next-generation accelerated computing.

Meta and NVIDIA have announced a collaboration to advance open AI networking with the introduction of Spectrum-X Ethernet SuperNIC and Spectrum-6 innovations. The partnership combines Meta’s experience operating AI infrastructure at hyperscale with NVIDIA’s networking technology, enabling agentic AI factories to scale beyond 100,000 GPUs with fewer switch tiers and lower infrastructure costs.

The announcement marks a significant step in developing an architecture and operational playbook that other AI infrastructure operators can adopt. This collaboration will help turn lessons from demanding production workloads into more scalable, resilient, and efficient AI factories.

NVIDIA and Cisco have announced plans for the Ultra-Dense Small Form Factor Pluggable (USFP) standard, which provides a new optical pluggable transceiver module form factor designed to scale next-generation AI fabrics while improving bandwidth density and power efficiency. The USFP will deliver four times the bandwidth density of traditional 200Gb/s-per-lane pluggable optics.

The result is the ability to connect more compute resources without adding network tiers that can increase latency, power consumption, and complexity. NVIDIA and Cisco are collaborating with TE Connectivity and Amphenol on cages, connectors, and modules to advance USFP through a community-led effort.

Google has announced Falcon Transport support on future NVIDIA ConnectX-10 SuperNICs and BlueField DPUs, helping improve network efficiency for large-scale AI infrastructure. The collaboration combines NVIDIA hardware acceleration with DOCA software to implement Falcon and extend NVIDIA networking to Google’s data center transport.

Falcon is a transport protocol designed to deliver data reliably over Ethernet networks by distributing traffic across paths, adapting to congestion, recovering dropped packets, and protecting connections. This helps workloads make better use of available bandwidth and maintain predictable performance as Ethernet fabrics scale.

Read more →

Building Context-Aware AI Assistants with AgentCore and OpenClaw

A new approach to building personal AI assistants has emerged, one that prioritizes continuity over individual question-answering. Traditional off-the-shelf AI assistants excel at answering specific questions but struggle to retain context between conversations. This limitation can be frustrating for users who need to re-explain their situation or preferences each time they interact with the assistant.

The problem lies not in the quality of the answers themselves, but rather that the assistant lacks memory and fails to accumulate context over time. To address this issue, a team has developed a personal AI assistant using OpenClaw, an open-source agentic system, running on AgentCore runtime, a capability of Amazon Bedrock AgentCore. This setup enables the creation of durable knowledge from disposable chats.

The architecture is domain-agnostic and can be applied to various domains such as support bots, fitness coaches, or internal help desks. The entire system lives within a single AWS CloudFormation template, which deploys with one command and runs on a consumption-based model that costs only a few dollars per month for light personal use.

To build this assistant, you’ll need Amazon Bedrock AgentCore access, including the runtime and memory capabilities. You’ll also require model access granted for the models you plan to route to, such as Claude Haiku 4.5 for text or Claude Sonnet 4.5 for vision. Additionally, you’ll need Docker with linux/arm64 build support and the AWS Command Line Interface (AWS CLI) configured.

The architecture consists of several key components: AgentCore runtime, OpenClaw as the agent substrate, two models routed by task, skills as reusable capability units, Telegram as the serverless front door, memory turning disposable chats into durable knowledge, and namespaces isolating each user’s data.

Read more →