Artificial intelligence agents are no longer a futuristic experiment confined to prototypes and pilot projects. Organizations have rapidly progressed from merely building AI agents to integrating them deeply into everyday workflows that span individuals, teams, and entire enterprises. However, the challenge has shifted: it’s no longer about crafting agents but about efficiently distributing them at scale, governing them securely, and embedding them naturally in the tools employees use daily.

Microsoft Foundry addresses this last-mile problem by enabling developers to publish agents directly into Microsoft 365 Copilot and Teams, deploy autonomous autopilot agents that collaborate across teams, and facilitate inter-agent communication seamlessly through Agent-to-Agent (A2A) protocols. This blog post explores these new capabilities, their architectural underpinnings, and how enterprises can unlock the full productivity potential of AI agents working alongside humans — all while maintaining strict governance and identity management.

We’ll start with an architectural overview that highlights how Foundry agents fit into the broader enterprise ecosystem, then dive into the key technical innovations in agent distribution and collaboration, followed by practical insights on how these components operate under the hood. Finally, we close with quick tips for developers and IT professionals looking to leverage Microsoft Foundry’s new enterprise agent features today.

Architecture Overview

(This section presents a conceptual architecture diagram illustrating Foundry’s integration with enterprise data, Microsoft 365 productivity tools, identity management via Microsoft Entra ID, and deployment into Microsoft 365 Copilot and Teams.)

This architecture reflects Foundry’s seamless integration of AI agents into existing enterprise data stores and productivity tools, governed centrally yet deployed organically within collaboration platforms like Teams. Agents operate with assigned identities and permissions, ensuring an enterprise-grade security model around automation.

Key Technical Observations

• Unified Publishing Pipeline to Microsoft 365 & Teams — Foundry’s consistent, governed publishing pipeline lets agents created once flow seamlessly into popular enterprise hubs without needing separate development efforts for each platform.

• Delegation-Centric Interaction Model — Moving beyond simple prompt/response chatbots, agents support goal-oriented delegation with continuous execution, checkpoints, and human collaboration natively through Teams and Microsoft 365 channels.

• Autonomous Autopilot Agents with Full User Identities — Autopilot agents have their own Microsoft 365 accounts, including email, calendar, OneDrive, and Teams access, allowing them to coordinate complex work streams residing in shared spaces rather than acting as isolated assistants.

• Agent-to-Agent (A2A) Interoperability — Built on an open protocol and enforced via Microsoft Entra ID authentication, A2A enables agents across varying frameworks and clouds to securely invoke each other, promoting modular and composable AI solutions.

• Enterprise-Grade Governance by Design — Agents are provisioned with Entra identities from creation; IT can manage them centrally through governance portals, ensuring that security and policy controls envelop every stage — from development through org-wide deployment.

• Context-Rich, Collaborative Workstream Management — The Workstream Manager autopilot agent demonstrates how agents can leverage chat history, files, and shared repositories dynamically, facilitating team coordination, tracking commitments, and surfacing risks.

How It Works

Publishing to Microsoft 365 Copilot and Teams

Developers create agents within Foundry’s agent factory and, when ready, submit them through a governed publishing process. This pipeline enforces identity registration with Microsoft Entra, access approvals, and deployment controls to ensure secure, enterprise-compliant rollout.

Once published, agents are accessible to employees inside Microsoft 365 Copilot and Teams without requiring users to learn new interfaces or jump between disparate apps. Instead of a simple chatbot prompt/response, users interact by assigning goals that the agent executes over time — providing progress updates, requesting approvals, and escalating issues. This model treats agents as autonomous collaborators rather than reactive tools.

# Example CLI snippet to publish an agent (conceptual)
foundry-agent publish --target teams,m365 --governed

Autopilot Agents: Autonomous Workstream Collaboration

Autopilot agents represent a new class of AI agents with independent Microsoft 365 user identities and access rights. This autonomy enables them to act like virtual team members:

• Reading and synthesizing conversation histories, meeting transcripts, and shared files
• Tracking open tasks, marking them visually in chats (📌 pin icons, reactions)
• Summarizing ongoing workstreams on demand with commands like /workstreamsummary
• Managing access control in group chats at the manager's discretion
• Leveraging Microsoft 365 services like Word, Excel, and Outlook for richer productivity, e.g., drafting emails or summarizing specifications with direct source access

This design acknowledges that enterprise work is a social, distributed effort crossing multiple contexts and roles.

Agent to Agent (A2A) Communication

Recognizing that complex workflows often span multiple specialized agents, Foundry supports inter-agent calls through an open A2A protocol:

• Outgoing A2A: Agents invoke remote agent endpoints as tools in their pipeline
• Incoming A2A (new): Agents can be exposed as A2A endpoints discovered via agent cards and invoked by any compatible client, framework, or platform

Microsoft Entra ID ensures all A2A calls are authenticated and authorized, preventing anonymous or malicious interactions. This interoperability layer fosters composability in AI ecosystems, allowing agents to call on each other without custom integrations or rewrites.

Governance and Identity Management

Every agent acquires an Entra identity upon instantiation, similar to a user or an application. These identities integrate with centralized management portals for:

• Reviewing and approving agent access and publishing requests
• Auditing agent use and invocation
• Assigning scopes and permissions based on business policies

Governance is embedded early, not retrofitted, aligning security and compliance workflows with AI automation agility.

Quick Tips & Tricks

1. Leverage Shared Spaces for Autopilot Agents — Design agents to work in Teams channels or group chats rather than just 1:1 chats. This enhances their organizational impact by sharing a single source of truth and contextual knowledge.

2. Use Built-in WorkIQ Integrations — Tap Foundry’s Microsoft 365 connectors to interface directly with Word, Excel, Outlook, and SharePoint to provide richer, source-accurate agent responses.

3. Implement Access Control Commands for Group Chats — Protect privacy and relevance by enabling manager-controlled user access to agents with commands like /access add and /access remove.

4. Adopt the Delegation Interaction Model — Shift your agent design from purely reactive Q&A to goal-driven tasks with checkpoints and milestone updates, improving usability and collaboration.

5. Expose Agents as A2A Endpoints — Develop agents that can be invoked by others via A2A to modularize complex workflows and leverage cross-platform AI toolchains.

6. Use Governance Portals for Lifecycle Management — Integrate agent approval and lifecycle management into your existing IT security operations for consistent compliance.

Conclusion

The evolution of Microsoft Foundry marks a pivotal moment where AI agents transition from isolated experiments into fully-fledged enterprise collaborators. By embedding agents directly into Microsoft 365 Copilot and Teams, assigning autonomous identities, enabling inter-agent communication, and enforcing governance from day one, Foundry empowers organizations to realize scalable AI-driven productivity across teams and workflows.

This new paradigm — AI not just assisting individuals but collaborating actively within organizations — promises to reimagine how work gets done. The open A2A protocol and integrated identity model set a foundation for interoperable, secure, and governed AI ecosystems that will expand as agent technologies mature.

For developers and enterprises alike, Microsoft Foundry offers a rich platform to build, publish, and operate AI agents that integrate naturally where employees spend their time, helping turn innovation into operational excellence.

References

1. From Building Agents to Working with Them: Enterprise Agent Distribution in Microsoft Foundry | Microsoft Foundry Blog — Official announcement and in-depth technical details on Foundry’s enterprise agent distribution roadmap.