Harness-agnostic how to delegate: split work, assign, isolate, verify, clean up.
- Owning workflows decide task goals; this skill defines how to delegate.
- Use whatever safe subagent mechanism the host provides.
- Host launcher/adapters own capability checks, launch, continuation, status, stop, and runtime-state cleanup — not this policy.
- Apply this policy with every launcher.
Delegate + split
When a safe launcher exists, delegate. “Small/easy” is not a reason to keep work in parent.
Parent keeps (strong reason only): framing/decomposition/synthesis · shared work document + ADR status · integrate/dispositions/acceptance · user comms · workspace + runtime cleanup · atomic no-judgment steps · explicit user prohibition
If you skip delegation, record why.
Split before launch:
- One independently verifiable outcome per lane
- Classify: reader · writer · independent validator
- Record scope, deps, expected output, join point
- Parallelize only independent lanes; coupled work → one awaited lane
- Cap fan-out by parent integrate/verify/cleanup capacity
- Prefer roles in
assets/agents/: scout · research · worker (adapt names to the launcher if needed)
Pick a launcher
Choose one accountable launcher per lane. Prefer one launcher for the whole workflow.
A launcher is safe only if it provides: exact cwd/context · permissions/write boundary · observable start/completion · output capture · timeout/bound · cancel when async · known cleanup ownership.
Candidates (first safe fit):
- Built-in / native subagent tools
- Host extension or plugin
- Catalog/runtime adapter skill for this harness
- Non-interactive agent CLI with the capabilities above
- Else: parent may execute; block only when delegation/independence itself is required
Never let two launchers control the same lane. Follow the chosen launcher’s mechanics for launch/status/stop/runtime cleanup.
Assignment contract
Build from assets/assignment-template.md + one profile in assets/agents/.
Each child gets one bounded job:
- Role + objective — one job, concrete outcome
- Mode — reader or writer
- Context — exact cwd/workspace + baseline, files, facts, open questions
- Scope — owned areas, requirements, non-goals, join output
- Permissions — least privilege; no secrets/prod/destructive acts unless explicit; no edits to the shared work document or ADR status; reviewers never edit their targets
- VCS — no child worktree/branch/commit/integrate/cleanup; read-only VCS only if authorized
- Validation — required checks; report exact results + skips
- Stop — completion condition, timeout; no recursive delegation
- Handoff — files read/changed · decisions · checks+results · skips · risks · blockers · remaining · evidence pointers (not transcripts)
Send only task-relevant context. Never send secrets, tokens, .env, or private transcripts.
Isolate (parent owns)
Unless the launcher explicitly owns isolation (parent still verifies):
- Readers: read-only. Shared checkout only if no concurrent writer can change the view and read-only is enforced; else isolate/snapshot
- Writers: sequential in one checkout after prior lane is stopped + dispositioned; concurrent only in isolated worktrees/workspaces with non-overlapping ownership
- Do not mutate an active writer cwd from parent/siblings
- Record before launch: baseline · path · branch (if any) · owner · existing dirty/untracked · deps · join point
Children never create/manage worktrees/branches or mutate VCS. Parent owns integrate + workspace cleanup.
Supervise
- Fresh child for new or independent judgment; reuse only for same-assignment follow-up
- Confirm start; monitor to completed / failed / cancelled / timed out / unknown
- Never fire-and-forget async work
- Repeated identical failure → change approach, don’t blind-retry
Verify + integrate
- Reach completed/stopped (or cancel) via the launcher
- Inspect handoff + full lane diff vs baseline (incl. untracked); child claims = evidence, not proof
- Rerun relevant checks in parent
- Integrate only if verified, via the repo’s normal method
- Keep the lane until join/acceptance checks pass
Cleanup (mandatory)
After each lane and at workflow end / interrupt recovery:
- Stop/cancel live children when the launcher allows
- Integrate, retain, or mark disposable workflow-owned work only
- Remove only safe resources: workflow-created (or launcher-owned), terminal, fully handled, not dirty/conflicted/unknown/user-owned, no pending join
- Clean launcher runtime/process state after review (preserve evidence until then) - do NOT leave dangling worktrees/branches/processes/sockets!
- Retain + report everything unsafe/unknown and why
No unaccounted workflow-owned resources.
Report
- roles / lanes used
- material results/changes
- parent verification
- unresolved/failed work
- worktrees/branches/workspaces created, integrated, removed
- retained resources + why
- launcher IDs only if needed for continuation