Coordination patterns for multi-agent GitHub delivery.
Choosing a coordination pattern
| Pattern | Use when | Cost |
|---|---|---|
| Sequential | Each step needs the previous step's output | Sum of steps |
| Parallel + barrier | You need all results together — dedup across findings, an early exit on zero | Slowest step |
| Pipeline (no barrier) | Each item flows through all stages independently | Slowest single chain |
| Hierarchical | A coordinator delegates and synthesizes | Slowest worker + synthesis |
| Adaptive loop | Repeat until a convergence condition (drive-to-green) | Unbounded — always cap rounds |
Default to pipeline over barrier. A barrier is justified only when a stage genuinely needs cross-item context. "I need to flatten the array first" is not a justification — do the transform inside a stage.
What parallelizes in GitHub work
Independent, run concurrently:
- Review lenses over the same diff
- Security scans across alert classes
- Metric collection across dimensions
- Per-PR triage across open PRs
Genuinely sequential:
- Branch → implement → PR (each needs the last)
- Landing a stack (each landing changes the next base)
- Migration → code → cleanup
The blackboard
Agents write findings to a shared structure; the coordinator reads and decides.
- Contradiction is information. Two agents disagreeing means one has context the other lacks. Investigate; do not average their confidence.
- Deduplicate by root cause, not by text. Two lenses describing the same defect differently is one finding.
- Record what was dropped and why. A board that reports only survivors is indistinguishable from a board that found nothing.
Delegation
- Narrow mandate plus an explicit return contract. "Review this PR" produces mush; "find correctness defects, return file/line/failure-scenario" produces findings.
- Keep subagent prompts under ~400 words — subagents inherit a large context chain before your prompt, and long prompts raise rejection rates.
- Prefer a named specialist over a general-purpose agent.
- Never spawn duplicate agents hoping one succeeds. Spawn them with different lenses so disagreement carries signal.
Merge gates
These are absolute. An agent may not route around them:
- Every required check green at the current head sha
- Review quorum met, no outstanding
CHANGES_REQUESTED - No unresolved threads when protection requires resolution
- Clean mergeability against the base
- No branch-protection bypass or admin override
- Human approval for
humanReviewPaths
Re-verify against live state immediately before merging. A gate that passed ten minutes ago is not evidence about now.
Reporting
Report per phase, with a marker per phase, and never claim a phase you skipped. A summary that reads complete while a step was silently dropped is the worst possible output — it converts a partial result into a false one.
See also
merge-queue— queue and protection semanticsreview-protocols— the adversarial board../workflows/README.md— declarative workflow authoring