Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Spawns multiple AI coding agents to work on related GitHub issues concurrently using git worktrees. Use when breaking down a large feature into multiple issues, running parallel agents with --print flag, or managing wave-based execution of related tasks.
.claude/skills/aiskillstore-orchestrating-parallel-agents/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 235% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 6% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -42% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 73% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -2% | 0% |
Spawn multiple Claude agents to work on related issues concurrently using git worktrees.
Copy and track progress:
Parallel Agent Orchestration:
- [ ] 1. Break feature into issues (1-3 files each)
- [ ] 2. Organize into waves (independent → dependent)
- [ ] 3. Pre-approve git permissions in settings.local.json
- [ ] 4. Spawn wave with --print flag
- [ ] 5. Monitor progress
- [ ] 6. Complete stragglers manually
- [ ] 7. Merge PRs (rebase between same-file conflicts)
- [ ] 8. Cleanup worktreesEach issue should be completable in isolation:
markdown## Problem What's broken or missing. ## Solution High-level approach. ## Files to Modify - `path/to/file` - what changes ## Implementation Code snippets or pseudocode. ## Acceptance Criteria - [ ] Testable outcomes
Key: Include file paths and code examples. Agents work best with concrete starting points.
Wave 1: Independent changes (no shared files)
Wave 2: Changes that may touch same files (expect conflicts)
Wave 3: Integration/testing (depends on all above)Rule: Same-file issues go in different waves OR same agent.
Add to .claude/settings.local.json for non-interactive --print mode:
json"Bash(git -C /absolute/path/to/worktree add:*)", "Bash(git -C /absolute/path/to/worktree commit:*)", "Bash(git -C /absolute/path/to/worktree push:*)"
bashfor issue in 101 102 103; do (claude --print "/worktree-issue $issue" > "issue-${issue}.log" 2>&1) & done
bashps aux | grep "claude.*worktree" | wc -l # Running agents git worktree list # Worktrees created tail -f issue-*.log # Live logs
If agent finishes code but fails on git:
bashgit -C <worktree> add -A git -C <worktree> commit -m "feat: description" git -C <worktree> push -u origin <branch> gh pr create --head <branch> --title "..." --body "Closes #N"
bashgh pr merge N --squash --delete-branch
If conflicts after prior merges:
bashcd <worktree> && git fetch origin main && git rebase origin/main # resolve conflicts git push --force-with-lease
bashgit worktree remove <path> git branch -D <branch> git worktree prune
| Tip | Why | |-----|-----| | 1-3 files per issue | Higher success rate | | Include "Files to Modify" | Agents find code faster | | Backend-first waves | Fewer frontend conflicts | | Merge same-file PRs sequentially | Rebase between each |
| Problem | Solution | |---------|----------| | Agent stuck on permissions | Complete git manually | | Merge conflict | Rebase, resolve, force-push | | Agent went off-scope | Reject PR, clarify issue | | Too many conflicts | Smaller waves, sequential merge |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 48,519 | 16,794 | -65% | 1 | 1 | 0% | 1,163 | 3,897 | +235% | 0 | 0 | — |
case-02 | fail→fail | 25,504 | 11,834 | -54% | 1 | 1 | 0% | 2,790 | 2,839 | +2% | 0 | 0 | — |
case-03 | fail→pass | 18,053 | 16,192 | -10% | 1 | 1 | 0% | 3,054 | 3,230 | +6% | 0 | 0 | — |
case-04 | pass→pass | 14,203 | 8,614 | -39% | 1 | 1 | 0% | 2,191 | 2,222 | +1% | 0 | 0 | — |
case-05 | fail→pass | 12,294 | 1,599 | -87% | 1 | 1 | 0% | 1,906 | 1,112 | -42% | 0 | 0 | — |
case-06 | pass→pass | 8,563 | 2,443 | -71% | 1 | 1 | 0% | 1,461 | 1,276 | -13% | 0 | 0 | — |
case-07 | pass→pass | 12,966 | 6,007 | -54% | 1 | 1 | 0% | 2,085 | 1,912 | -8% | 0 | 0 | — |
case-08 | pass→pass | 8,341 | 3,480 | -58% | 1 | 1 | 0% | 1,316 | 1,448 | +10% | 0 | 0 | — |
case-09 | fail→pass | 5,142 | 2,994 | -42% | 1 | 1 | 0% | 810 | 1,400 | +73% | 0 | 0 | — |
case-10 | pass→pass | 9,884 | 3,031 | -69% | 1 | 1 | 0% | 1,607 | 1,329 | -17% | 0 | 0 | — |
case-11 | pass→pass | 12,055 | 4,869 | -60% | 1 | 1 | 0% | 1,996 | 1,678 | -16% | 0 | 0 | — |
case-12 | pass→pass | 11,454 | 4,159 | -64% | 1 | 1 | 0% | 1,667 | 1,457 | -13% | 0 | 0 | — |
case-13 | fail→pass | 12,682 | 5,924 | -53% | 1 | 1 | 0% | 2,071 | 2,024 | -2% | 0 | 0 | — |
case-14 | pass→pass | 13,212 | 8,493 | -36% | 1 | 1 | 0% | 2,062 | 2,216 | +7% | 0 | 0 | — |
case-15 | pass→pass | 11,948 | 3,600 | -70% | 1 | 1 | 0% | 1,653 | 1,471 | -11% | 0 | 0 | — |
case-16 | fail→pass | 7,131 | 2,423 | -66% | 1 | 1 | 0% | 1,088 | 1,264 | +16% | 0 | 0 | — |
case-17 | fail→pass | 15,777 | 14,109 | -11% | 1 | 1 | 0% | 2,636 | 3,049 | +16% | 0 | 0 | — |
case-18 | pass→pass | 5,728 | 1,760 | -69% | 1 | 1 | 0% | 972 | 1,066 | +10% | 0 | 0 | — |
case-19 | fail→pass | 14,613 | 5,425 | -63% | 1 | 1 | 0% | 2,191 | 1,865 | -15% | 0 | 0 | — |
case-20 | pass→pass | 16,284 | 18,793 | +15% | 1 | 1 | 0% | 2,962 | 4,170 | +41% | 0 | 0 | — |
case-21 | pass→pass | 14,618 | 9,844 | -33% | 1 | 1 | 0% | 2,326 | 2,765 | +19% | 0 | 0 | — |
case-22 | pass→pass | 9,962 | 7,600 | -24% | 1 | 1 | 0% | 1,985 | 2,319 | +17% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.