multi-agent-handoff

Multi-agent handoff

Structure work so multiple agents can execute sequentially without losing context.

When to use

• Splitting a large task across explorer, implementer, and reviewer agents
• Foreman/worker or parent/subagent patterns
• Need a written baton between chat sessions or tools

One obligation

Name one outcome. After two attempts on the same owner with no new evidence, hand off with what was tried, the evidence, and the non-claims. Adjacent discoveries are part of the baton, not permission for the sender to absorb the next problem. MoE findings do not authorize writes.

A handoff does not create an owner. Skill Steward does not schedule claims or presence. Do not add more baton fields. See work discipline.

Handoff document template

Create or update HANDOFF.md (or a section in the task issue) with:

## Goal
{one sentence outcome}

## Done
- {completed items}

## Next
1. {ordered steps for the receiving agent}

## Constraints
- {tech stack, style, files not to touch}

## Verification
- {commands or checks that must pass}

## Validation status
- {commands run}
- {commands skipped or blocked, with reason}
- {blocked JSON explained with `steward blocked explain --input <path> --json`, when available}
- {schema/output drift checked with `steward schema check-outputs --json`, when machine-readable output is part of the handoff}
- {claims not proven because validation was skipped or blocked}

## Partial results
- {missing, partial, superseded, or timed-out agents/lenses}

## Context links
- {paths, PRs, prior decisions}

## Artifact capture
- {ADR, FAQ, skill, evidence note, test, validator, generator, or check that should absorb durable learning}

Parallel batch contract

For broad decomposable work, the parent may use a disposable batch section instead of a new plan format. Keep only enough contract to move safely:

• original goal and user acceptance check;
• default native gate and aggregate gates;
• product impact check for the primary artifact, especially when the repo is an app, library, CLI/tool, plugin, or prototype;
• detour budget and stop condition;
• integration capacity, merge order, and conflict policy;
• comparison strategy for lane outputs;
• final evidence boundary, claim ceiling, and non-claims;
• acceleration note with three fields: Saved, Cost/duplication, and Future hot path;
• hot-path promotion check for repeated verification or comparison work.