{"id":46694,"date":"2025-09-24T08:00:00","date_gmt":"2025-09-24T15:00:00","guid":{"rendered":""},"modified":"2026-04-07T11:42:19","modified_gmt":"2026-04-07T18:42:19","slug":"agent-factory-designing-the-open-agentic-web-stack","status":"publish","type":"post","link":"https:\/\/azure.microsoft.com\/en-us\/blog\/agent-factory-designing-the-open-agentic-web-stack\/","title":{"rendered":"Agent Factory: Designing the open agentic web stack"},"content":{"rendered":"\n

This blog is a wrap-up post in a blog series called Agent Factory<\/a> which shares best practices, design patterns, and tools to help guide you through adopting and building agentic AI.<\/em><\/p>\n\n\n\n

The rise of AI agents\u2014autonomous software entities acting on behalf of users and organizations\u2014marks a transformative moment for enterprise technology. But as we\u2019ve explored throughout this blog series, building effective agents is about more than just code. It requires a repeatable blueprint, spanning use case design, developer tooling, observability, integrations, and governance. <\/p>\n\n\n\n

Throughout this series<\/a>, we\u2019ve walked through the journey of building enterprise-grade agents: from early use cases and design patterns<\/strong> to the tools and developer workflows<\/strong> needed to move from prototype to production, to the importance of observability, interoperability, and open standards<\/strong>, and finally the governance and security principles<\/strong> required to deploy agents responsibly. <\/p>\n\n\n\n

Now, as we conclude the series, we zoom out to the bigger picture<\/strong>: the agentic web stack<\/em>. Much like HTTP and TCP\/IP standardized the internet, this stack provides the common services and protocols needed to make multi-agent ecosystems secure, scalable, and interoperable<\/strong> across organizational boundaries. <\/p>\n\n\n\n

\n
Learn more with Azure AI Foundry<\/a><\/div>\n<\/div>\n\n\n
\"Diagram<\/figure>\n\n\n\n

Blueprint: 8 essential components and services<\/h2>\n\n\n\n

A robust agentic web stack is not one technology but a composition of services<\/strong> that together provide the foundation for open, secure, and enterprise-grade multi-agent systems. Here\u2019s what it takes\u2014and how Azure AI Foundry<\/a> is making it real. <\/p>\n\n\n\n

1. Communication protocol service<\/h3>\n\n\n\n

Agents need a shared \u201clanguage\u201d to exchange messages, requests, and structured data. Without it, collaboration breaks down into isolated silos. Standards like Model Context Protocol (MCP) and<\/strong> Agent-to-Agent (A2A)<\/strong> provide this foundation, ensuring agents can negotiate, coordinate, and cooperate\u2014regardless of who built them or where they\u2019re hosted. In Azure AI Foundry, A2A support<\/a> enables not only intra-organization workflows but also cross-boundary collaboration, where supply chain partners or business ecosystems can securely exchange actions through a common protocol.<\/p>\n\n\n\n

2. Discovery registry service<\/h3>\n\n\n\n

Just as the web needed directories and search engines, agents need a way to be found and reused. The Catalog<\/strong> serves as the listing of assets\u2014a curated collection of agents, tools, and applications that can be discovered and composed into new solutions. The Registry<\/strong>, by contrast, tracks the deployed instances of those assets\u2014the live agentic app instances running across providers, with their endpoints, health, and status. Together, the Catalog and Registry bridge the gap between what\u2019s available and what\u2019s active, enabling developers to publish once, discover broadly, and reliably orchestrate workflows against running systems.<\/p>\n\n\n\n

3. Identity and trust management service<\/h3>\n\n\n\n

Trust is the lifeblood of the agentic web. Every agent must carry a verifiable identity, enforced through standards like OIDC (OpenID Connect) and JWT (JSON Web Token)<\/strong>, and tied into enterprise systems like Microsoft Entra ID<\/a><\/strong>. In Azure AI Foundry, identity is not an afterthought\u2014it\u2019s the control plane. This enables fine-grained role-based access, ensures that only authorized actors participate in workflows, and provides auditable accountability for every action an agent takes. Combined with encrypted channels, this identity-first model enforces a zero-trust security posture<\/strong> across the agentic stack.<\/p>\n\n\n\n

4. Tool invocation and integration service<\/h3>\n\n\n\n

No agent can succeed in isolation; value comes when agents can connect with data, APIs, and enterprise systems. The Model Context Protocol (MCP)<\/a><\/strong> provides a vendor-neutral standard for exposing tools in a way that any compliant agent can invoke. In Azure AI Foundry, MCP integration is deeply embedded, so developers can register enterprise APIs once and instantly make them available to multiple agents\u2014whether they\u2019re built on Microsoft\u2019s agent frameworks like Semantic Kernel and AutoGen, LangGraph, or third-party SDKs. This eliminates bespoke integrations and allows enterprises to compose workflows from best-in-class components.<\/p>\n\n\n\n

5. Orchestration service<\/h3>\n\n\n\n

Single agents can handle discrete tasks, but the real breakthroughs come from multi-agent orchestration<\/strong>: teams of agents collaborating across multi-step, distributed processes. Azure AI Foundry delivers this through a unified framework<\/a> that brings together Semantic Kernel and AutoGen\u2014and extends it with multi-agent workflow orchestration<\/strong> inside Azure AI Foundry Agent Service. These workflows manage dependencies, allocate resources, and resolve conflicts, enabling enterprise-scale use cases such as financial transaction processing or IT incident response.<\/p>\n\n\n\n

6. Telemetry and observability service<\/h3>\n\n\n\n

As we covered in Part 3<\/a>, observability is non-negotiable for reliable agents. Azure AI Foundry extends OpenTelemetry<\/strong> with agent-aware instrumentation\u2014tracing conversations, capturing performance data, and surfacing anomalies in real time. This makes agent behavior explainable and debuggable, while also serving governance needs: every decision and action is logged, auditable, and tied back to identity. For enterprises, this is the bedrock of trust, compliance, and continuous improvement<\/strong>.<\/p>\n\n\n\n

7. Memory service<\/h3>\n\n\n\n

Agents without memory are limited to stateless interactions; agents with memory become adaptive, contextual, and human-like in their continuity. Azure AI Foundry supports both short-term session memory and long-term enterprise knowledge integration. Imagine a customer support agent that recalls prior interactions across channels, or a supply chain agent that tracks historical disruptions to improve future decisions. With memory, agents evolve from transactional helpers into strategic partners that learn and adapt over time<\/strong>.<\/p>\n\n\n\n

8. Evaluation and governance service<\/h3>\n\n\n\n

Finally, no stack is complete without governance<\/strong><\/a>. This includes continuous evaluation, policy enforcement, ethical safeguards, and regulatory compliance. In Azure AI Foundry, governance hooks are built into orchestration, observability, and identity services\u2014enabling enterprises to block unsafe actions, enforce approvals for sensitive workflows, and generate compliance-ready audit trails. This ensures organizations don\u2019t just innovate fast, but innovate responsibly<\/strong>.<\/p>\n\n\n\n

Strategic use cases and business value<\/h2>\n\n\n\n

The agentic web stack is not theoretical; it unlocks concrete enterprise value.<\/p>\n\n\n\n