Reliable Multi-Agent Systems: Orchestration, State, and Human Approval

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