subagent-skill-harvest-loop

name	subagent_skill_harvest_loop
description	Use when running delegated subagent rounds to generate prototypes, harvest evidence and feedback, patch skills with extracted knowledge, and clean up temporary project workspaces.

subagent_skill_harvest_loop

Quick Start

1. Define the round objective and target skill(s) to improve.
2. Create isolated per-agent work directories.
3. Spawn subagents with disjoint ownership and strict deliverables.
4. Require each agent to produce prototype + feedback + DOC_PACK artifacts.
5. Harvest and synthesize cross-agent findings.
6. Patch skills/references with extracted knowledge.
7. Sync updated skills globally and verify parity.
8. Delete temporary round workspaces and close agents.

When To Use

• You want subagents to explore a task family and produce evidence for skill improvements.
• You need repeatable delegation + synthesis + cleanup cycles.
• You want reusable templates for assignment quality and artifact extraction.

Inputs Required

• Objective statement for the round.
• Number of agents and disjoint ownership paths.
• Mandatory artifact schema.
• Skill targets to patch.
• Cleanup policy.

Workflow

1. Draft assignments using references/delegation-assignment-template.md.
2. Enforce artifact schema from references/subagent-feedback-schema.md.
3. Collect results with references/harvest-synthesis-matrix.md.
4. Patch skill docs via references/skill-patch-checklist.md.
5. Run cleanup and verification via references/cleanup-and-verification-checklist.md.

Guardrails

• Never allow overlapping agent write ownership.
• Require explicit non-destructive cleanup rules.
• Do not patch skills from single-agent anecdotes; synthesize across artifacts.
• Keep source evidence paths in every final synthesis report.
• Always verify local/global skill parity after patching.

Output Contract

Return:

• delegation summary,
• harvested insights matrix,
• skill patches applied,
• global sync verification,
• cleanup confirmation.

Validation Checklist

• All agent outputs satisfy required artifact schema.
• Synthesized findings are cross-agent and evidence-backed.
• Skill patches are concrete and reference-linked.
• Temporary workspaces are removed.
• Agent sessions are closed.

Failure Modes + Recovery

• Missing artifacts: re-request targeted completion from specific agent.
• Inconsistent findings: prioritize repeated patterns across agents.
• Partial patching: produce patch backlog and apply in ranked order.
• Cleanup drift: run explicit directory and session closure checks.