Microsoft Reactor London

Adding more agents does not automatically create a better system. Multi-agent solutions need clear responsibilities, controlled execution, recoverable state, and carefully designed decision boundaries.

In this session, we will extend the agent built in Episode 1 into a coordinated incident-management workflow. Specialist agents will handle triage, knowledge retrieval, impact assessment, and remediation planning. We will use Microsoft Agent Framework to demonstrate sequential, concurrent, and handoff orchestration patterns, and discuss where each pattern is appropriate.

The workflow will introduce typed state, checkpointing, failure recovery, and human-in-the-loop approval before any sensitive remediation action can continue. We will also separate deterministic application logic from agentic reasoning, showing which decisions should be handled by an AI agent and which should remain enforced by conventional code.

Attendees will see how a multi-agent workflow behaves when interrupted, how it resumes from saved state, and how human approval can be incorporated without losing execution context.

By the end of the session, attendees will be able to design controlled multi-agent workflows that combine AI judgment with reliable business-process execution. Key topics: Multi-agent orchestration patterns Sequential, concurrent, and handoff workflows Typed workflow state Checkpointing and recovery Human-in-the-loop approvals Agentic versus deterministic steps Failure handling and idempotency.

An enterprise agent becomes genuinely useful when it can act on business systems. It also becomes significantly more dangerous if those integrations are not governed correctly.

In this session, we will connect our Microsoft Foundry agent to real enterprise capabilities using Azure API Management, Azure Logic Apps, Azure Functions, Azure Service Bus, OpenAPI tools, and Model Context Protocol servers.

The agent will create an incident ticket, initiate an approval workflow, invoke a protected business API, submit a long-running remediation command, and report the final result back to the user. Rather than allowing the model direct access to backend systems, we will introduce a governed integration layer with typed tool contracts, authentication, authorization, rate limiting, schema validation, retries, and centralized logging.

We will also compare the roles of OpenAPI tools, MCP servers, Logic Apps actions, Azure Functions, and asynchronous messaging. Managed identities and Microsoft Entra ID will be used to eliminate embedded API keys and apply least-privilege access across the solution.

By the end of the session, attendees will understand how to give agents meaningful enterprise capabilities without bypassing security, integration governance, or operational controls. Key topics: Microsoft Foundry tools OpenAPI and Model Context Protocol Azure API Management as an AI gateway Azure Logic Apps agent actions Azure Functions Azure Service Bus Managed identity and Microsoft Entra ID Secure and asynchronous tool execution.

--------------------------------------------------------------