Source: Handoff orchestration (Microsoft Agent Framework docs)

Agent Framework's mesh-topology handoff orchestration: handoff rules, context broadcast, interactive vs autonomous mode, and a stated handoff vs agent-as-tools contrast.

Contents

Summary

Handoff orchestration connects agents in a mesh with no central orchestrator; each agent decides, by calling an injected handoff tool, whether to pass the conversation on. The page contrasts this with agent-as-tools on control flow, task ownership and context, and explains how agents with separate sessions are kept in sync.

Key claims

  • Contrast: in a handoff the receiver takes full ownership and has the conversation so far; with agent-as-tools the primary agent keeps responsibility, gets control back and passes tool agents only what they need.
  • HandoffBuilder (Python) / AgentWorkflowBuilder.CreateHandoffBuilderWith (C#) register handoff tools automatically; add_handoff / WithHandoffs restrict who may hand off to whom. By default all agents can hand off to each other.
  • Agents do not share one session. Each broadcasts user and agent messages to all participants; tool content, including handoff calls and results, is filtered out before forwarding. Python keeps text only for non-user messages.
  • Handoff is interactive by default: an agent that answers without handing off triggers a request for user input. Experimental autonomous mode replies with a continuation prompt instead, up to 50 turns per agent by default.
  • Python handoff tool names derive from the target agent ID, sanitised and capped at 64 characters. When checkpointing, agent IDs and names must stay stable across rebuilds.
  • Approval-required tools pause the workflow; checkpoint storage lets approvals arrive much later.

Feeds

Suggest changes →