Agents visual guide

A verifiable multi-agent handoff

How scoped context and evidence move between an orchestrator, worker, independent reviewer, and human.

Editable sourceMermaidOpen full-size SVG ↗Download .mmd source

A sequence diagram shows a Human sending a goal and constraints to an Orchestrator. The orchestrator sends a scoped task and acceptance criteria to a Worker. The worker returns output, evidence, and open questions. The orchestrator sends the original criteria and output to an Independent reviewer, who returns pass or findings. Failures return to the worker for at most two revisions; a pass returns a decision packet to the human. — A trustworthy handoff preserves the original criteria, limits context, returns evidence, and uses an independent review before human acceptance.

Read the visual

What this shows

Parallel agents are useful only when their contracts are explicit. The orchestrator sends a bounded task with inputs and acceptance criteria, the worker returns outputs and evidence, an independent reviewer evaluates the result, and the human receives a concise decision packet with unresolved risks.

Carry this forward

Key takeaways

Every handoff needs scope, inputs, outputs, acceptance criteria, and stop conditions.
The reviewer needs the original task and evidence, not just the worker's summary.
Bound revision loops and escalate unresolved disagreements to a human.