Orchestrate Subagents

Orchestrate Subagents

Overview

Run coding subagents reliably: define a task bundle (often spanning several related issues), create isolated worktrees, spawn 1 or 3 identical attempts, then select the best PR and clean up the rest.

This skill is for the orchestrator/PM agent. Coding agents do implementation work and need not spawn subagents themselves.

Workflow (Best-Of-3 Default)

1) Write a bundle spec

• Put the bundle spec in /scratch/ (outside the repo), so you don’t create repo noise.
  • Example path: /scratch/claude-bundle-spec-<date>-<slug>.md
• Include:
  1. Issue list (IDs + titles).
  2. Acceptance criteria (bulleted, testable).
  3. Out of scope (explicit “DO NOT …” list).
  4. Required readings (exact file paths relative to worktree root).
  5. Definition of done (tests/checks to run, CI must be green).
  6. “Stop and ask” triggers (ambiguity, scope conflicts, failing tests that look unrelated).

2) Write the subagent prompt file

• Use one prompt file for all subagents (identical attempts):
  • Example path: /scratch/claude-prompt-<date>-<slug>.md
• Use an absolute prompt path in claude code exec (e.g. a /scratch/... path).
• Transform the bundle spec into a prompt. Add instructions that are about how the subagent has to work, beyond the bundle spec.
  1. Work only in their current worktree (pwd reveals it).
  2. Run onboarding/sanity: bash -lc scripts/hello.sh. Subagents read AGENTS.md and will know what to do.
  3. Implement the full issue bundle (single PR).
  4. Run the relevant checks (at minimum: npm run check).
  5. Open a PR (with a clear title and issue-closing footer).
  6. Stop and report back if they hit non-trivial blockers (don’t “invent” scope). Minimal prompt template (edit for the specific bundle):

# Prompt
You are a coding sub-agent. Work only inside your current git worktree.
## Onboarding
- Run `bash -lc scripts/hello.sh`.
## Task bundle
- Solve issues: #<A>, #<B>, #<C> as ONE coherent PR.
### Acceptance criteria:
- (list)
### Out of scope:
- (list explicit NOTs)
## Constraints
- Do not do PM tasks. Do implementation only.
- Do not add temporary/demo files.
## Process
- If you find ambiguity or a likely typo in the issue text, STOP and write a short question in the PR description.
## Validation
- (list)
## Deliverable
- Push a branch and open a PR.
- PR description must:
  - Reference all issues with “Closes #…”. 
  - Include a short, skimmable summary and any tradeoffs/risks.
  - Inform the reviewer of any deviations from the acceptance criteria.
  - Inform the reviewer of any shortcomings or known issues you did not address, and why you did not address them.

3) Create Worktrees and Spawn Subagents

• For best-of-3, create 3 worktrees with 1 subagent each.
• See .claude/skills/git-worktrees/SKILL.md for worktree management.
  • Create: scripts/worktree-new.sh <path> <branch> [--force]
  • Remove: scripts/worktree-remove.sh <path> [--force]
• Recommended naming (deterministic, easy to filter in PR list):
  • Slug: <yyyymmdd>-<short-bundle-name>
  • Branches: agent/<slug>/a, agent/<slug>/b, agent/<slug>/c
  • Paths: /workspaces/worktrees/<slug>-a (and -b, -c)
• Start each subagent from inside its worktree directory (no backgrounding):
  • cd /workspaces/worktrees/<slug>-a && claude code exec "Prompt: /scratch/claude-prompt-<date>-<slug>.md" 2>/dev/null

Important:

• Do not add & to the claude code exec command.
• To run best-of-3 in parallel, run one subagent per terminal (three terminals total).

4) Monitor and Triage

• We suppress stderr because it is extremely verbose and not meant for orchestrator review.
• When the agent is done, it exits and prints its final message to stdout. It also ought to have opened a PR as per the prompt instructions.
• Wait for all three agents to finish (process exit) before proceeding to selection. Agents maybe will not write to stdout at all until they are done.
• If a subagent fails early, e.g. due to misspecified scope or unclear instructions or violated task assumptions, you should usually escalate to the owner instead of trying to salvage the run.
• If a subagent never even starts, e.g. due to a typo in the claude code exec command, you can usually just restart it without escalating.

5) Select the winning PR (best-of-3)

• Once all three subagents have finished, review their PRs.
• Prefer the PR that:
  1. Meets the bundle acceptance criteria, including passing all tests.
  2. Has the highest quality in code architecture, code style, and test coverage.
  3. Is a solid foundation for the remaining project work.
• The PR text is useful commentary, but not a source of truth: the worktree contents are what will be merged and thus the only thing that matters.

6) Merge, Close, and Clean Up

• Merge the winning PR via GitHub CLI.
• If the merged PR fell short of acceptance criteria, create follow-up issues for the remaining work.
• Close losing PRs with a short note (“best-of-3; superseded by #”).
• Delete their remote branches.
• Pull in the main worktree so the merged result is reflected locally.
• Remove the winning and the losing local worktrees.
• Hand-off to the owner with a summary:
  1. Winner PR link + what it does.
  2. Which PRs were closed and why.
  3. Anything that needs owner review/decision.
  4. Follow-up issues suggested (if any).