agentleFS
p10ns11yp10ns11y/skills/agent-orchestrator/SKILL.md

agent-orchestrator

Skill2 starsChanged 40 days agoMITView on GitHub

Triage then orchestrate workers with briefs and verify-before-done. Use when coordinating, delegating, or reviewing agent output.

12 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use brainstorming.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • OpenAISkill

    brainstorming

    You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.

    Also in 3 projects that use this

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Not for one-file typo fixes (single_shot is the skip path).

    line 128
  2. Very specific:

    Does not replace domain skills.

    line 130

“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

agent-orchestrator

Load rule: Formal SoT below. Expand references/english-procedure.md only if a phase or handoff is still ambiguous. CLT: on-the-fly pre-filter ../rules/clt-dual-load.mdc (A8 DualLoad) — expand ../control-graph/references/clt-load-balance.md only if human/agent load tradeoff is ambiguous.

Purpose

Coordinate non-trivial work so workers implement and the orchestrator re-runs verify. Skip for one-file typo fixes.

Prerequisites

  • Target repo AGENTS.md for verify cmds.
  • Git worktrees for concurrent workers (git-worktrees).
  • Brief template: references/task-brief.md.

Instructions

Orch ≔ coordinator   // you
W    ≔ workers        // one brief, one workspace, one DoD each
Triage ∈ { single_shot, light, full }
Wave  ∈ { spike, foundation, feature, hygiene }

A1  Workers never self-verify as final truth — Orch re-runs verify cmds
A2  Brief before delegate; commit in workspace before "done"
A3  Merge ≻ cp   // integration via git only ([git-worktrees](../git-worktrees/SKILL.md))
A4  Spike hard/unknown BEFORE mega-feature brief
A5  Resume = gap-only brief  — never re-delegate full original scope
A6  Compose control shape with [control-graph](../control-graph/SKILL.md); this skill = multi-worker logistics

Pair: git-worktrees · concurrent-cli-agents · split-to-prs · domain skills per worker.

Triage (~10s)

ModeWhenDo
single_shot1 clear change · ≤1–2 files · low risk · obvious verify · no concurrent Wedit + verify; skip brief/worktree
lightplan/review only · 3–5 files same concern · tiny resume gap · fast disprove3–5 bullets then implement in one checkout
fullmulti W/worktrees · unknown hard part · large/vague · high stakes · prior agent "done" · resume after dayschecklist + briefs + verify + integrate
request → triage → single_shot | light | full
escalate: single_shot grows past ~2 files / new unknowns → stop → light|full

Full path

ORIENT → DECOMPOSE → SPIKE? → BRIEF → DELEGATE → VERIFY ⇄ fix-brief → INTEGRATE → REPORT
PhaseExit when
ORIENT1-sentence goal · base ref · constraints · unknowns · risk
DECOMPOSEwaves with disjoint ownership where possible
SPIKErisky unknown proved or rejected (≤1–2 files / readonly + rec)
BRIEFreferences/task-brief.md complete
DELEGATEone prompt/W; workspace recorded
VERIFYOrch ran every verify cmd; scope matches brief
INTEGRATEmerge one branch at a time; re-verify; no cp
REPORTsummary to user
RoleOwns
Orchspec, sequence, verify, merge order, user updates
Workerimplement in one worktree; commit on agent/<tool>/<slug>
Userscope splits, merge/PR, destructive git

Orch may implement only if trivial or verify failed twice and inline fix is faster — say so.

Brief minimum

  1. Problem · 2. Outcome (observable) · 3. Non-goals · 4. Standards (skills)
  2. Files may/must-not · 6. Verify cmds (exact) · 7. Workspace · 8. Done artifact (commit)

Worker handback: SHA · cmds+exit · 2-sentence approach · limits.

Verify (Orch-owned)

CheckAction
Scopegit diff base...agent ⊆ allowed
Claimssummary matches diff
Automatedre-run every brief command
Regressionimports/types/exports
Standardsspot-check cited skills

Fail → fix brief (gap only), same workspace if possible. Never merge on testimony alone.

Resume

read last brief → git log/diff/worktree list → remaining gap (1¶) → "Resume: …" brief only

Defaults (override from target AGENTS.md): frozen lockfile when deps change; type-check + lint; React → react-client-expert; deps → fix-dependency-security; minimal diff.

Stack handoff: leave a pushed linear branch + per-PR SHAs. Confirm the final branch name with the user. Detail: references/english-procedure.md.

Examples

User: "Split this refactor across two agents. Do not merge on their word."

Agent: triage full. Write two disjoint briefs with exact verify cmds. Each worker gets one worktree. Orch re-runs every verify cmd, then merges via git (never cp). Report SHAs + leftover gaps.

User: "Fix the typo in README."

Agent: triage single_shot. Edit + verify. No worktree, no brief.

Limitations

  • Not for one-file typo fixes (single_shot is the skip path).
  • Orch must not both implement and "verify" the same large feature.
  • Does not replace domain skills; workers load those.
  • Graphite/gt session of the agent may not be the user's authenticated shell.

Troubleshooting

Error / symptomCauseFix
Worker says done, tests failTestimony without Orch re-runRe-run brief verify cmds; gap-only fix brief
Diff outside allowed filesVague briefTighten files may/must-not; same workspace
Merge conflict stormOverlapping wavesDecompose disjoint; merge one branch at a time
Resume redoes the whole taskFull original brief reusedGap-only resume brief
cp from worktreeIntegration shortcutmerge / cherry-pick only

Report template

## Orchestrator summary
**Goal:** …
**Workers:** tool/branch → pass/fail
### Verified — commands + exit
### Integrated — merges
### Leftovers / next wave

Done when: triage mode chosen; if full — verify cmds passed on integration; user has report + leftovers.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.