| name | orchestration/multi-agent-orchestration |
| trigger-words | ["orchestration/multi","agent","orchestration"] |
| description | Orchestrate parallel AI agent workstreams — multi-agent deep research (3 parallel experts), research orchestration
Trigger phrases: orchestration/multi, agent, orchestration.
Converted from Hermes-Skill (Yuno's mastery library).
|
Multi-Agent Orchestration
Orchestrate parallel AI agent workstreams for research, audits, and development.
The Pattern
set -euo pipefail
Phase 1: Parallel Research (3 experts, ~5-10 min each)
Phase 2: Immediate Fixes (parent acts while reviewing results)
Phase 3: Synthesis (merge reports, prioritize) + 3.5 Cross-Check + 3.6 User-Decision
Phase 4: Execute & Document (P0 fixes first, write report to ~/docs/)
Phase 5: Retrospective (what worked, what didn't)
When to Use
- Deep research/improvement plans across multiple areas
- System audits requiring breadth + depth
- Any task splitting into 3+ independent domains
- Self-skill evaluation — scan and evaluate one's own skill library
- Meta-orchestration — analyze which orchestration patterns exist across skills
For recurring health checks, use no_agent: True shell scripts that exit silently on success. Only output when there's a problem.
Phase 1: Spawn 3 Experts
Use delegate_task(tasks=[...]) with 3 parallel tasks. Full briefing template, scope assignment, toolset table, and parent-direct fallback: see references/phase-1-spawn-experts.md.
Typical scopes (adapt to actual system): Expert 1 Auth/Security/Credentials, Expert 2 Tools/APIs/GPU-Compute, Expert 3 Infra/Messaging/Monitoring.
Critical: Experts without terminal + file produce estimates, not measurements.
Phase 2: Immediate Fixes
While experts run, execute obvious quick wins in the parent: re-enable deactivated toolsets, fix permissions, clear caches, verify subagent claims (don't trust "✅ done").
Phase 3: Synthesis
Deduplicate, conflict-check, verify subagent claims, identify blocking deps, prioritize P0/P1/P2/P3. Two proven sub-procedures are required (full detail in references/phase-3-synthesis.md):
- Phase 3.5 — Critical-Cross-Check: Grep-verify the 3-5 most-actionable claims per scout BEFORE proposing fixes. Pitfall #5 (verify EVERY claim).
- Phase 3.6 — User-Anker Decision: After cross-check, present 2-4 verified fixes + 1-2 deferred items. User picks, parent doesn't dump 5+ fixes at once.
Both sub-phases include the 3-Tier Verifikations-Matrix (Datei-Existenz + Content-Validierung + Realitäts-Check) and 4 Recovery Options for timed-out subagents.
Phase 4: Execute & Document
Execute P0 fixes first in dependency order. Write report to ~/docs/builds/ or ~/docs/system/. Update ~/docs/builds/README.md. Report template: references/report-template.md.
Phase 5: Retrospective
Key questions: (1) Were the 3 scopes correct? (2) Did any expert produce estimates instead of measurements? (3) Were CLI commands validated before execution? (4) What would a 4th "Validator" expert have caught?
Queen-Bee Orchestration
Die Bienenkönigin orchestriert einen Schwarm aus Subagenten. Das Delegation-Modell ist unabhängig vom Parent-Modell — die einzige Steuerung ist delegation.model in ~/.hermes/config.yaml (Pitfall #10: model-Param wird ignoriert).
Rollen: Königin (stark, höher Kosten, Strategie/Synthese) → Scouts (free tier, parallel, Messungen) → Arbeiter (Cron-Script, no_agent=True, recurring checks). Bei Free-Tier Rate-Limits → Parent übernimmt direkt (Fallback Phase 2).
Full config blocks, free-model recommendations (Owl Alpha, Nemotron, Qwen3 Coder), Token-Plan-aware multi-provider routing (Profile + MoA-Preset patterns), and anti-patterns: see references/queen-bee-configuration.md.
Batch vs Subagents Decision Matrix
| Criterion | Subagents | Python Batch |
|---|
| File count | 1-10 per expert | 20-1000+ |
| Check type | Needs reasoning | Deterministic |
| Context cost | High | Low |
| Risk | Phantom-fixes | Parent sees exact output |
Pitfalls (Quick Reference)
The full 20-pitfall detailed list with examples lives in references/pitfalls-detailed.md. For symptom-based trigger-watchlist (load BEFORE every delegate_task): see companion skill multi-agent-pitfalls-cheatsheet.
Top 10 one-liners (must defend against):
--help first — never assume CLI syntax- Timeouts 60-120s+ — network/build/install
- terminal+file toolsets — otherwise estimates not measurements
- Pre-check deps — "5 min" estimates are optimistic
- VERIFY EVERY CLAIM — subagent self-reports are NOT facts
- Explicit OUTPUT path —
~/docs/system/NAME-DATE.md, never defaults
- Parent applies config — subagents blocked on
.env/config.yaml writes
- Parent-only YAML edits — never let multiple subagents touch same config
- MAX 8 web-calls — subagents WILL loop pagination otherwise
model param ignored — only delegation.model config controls routing
Critical advanced pitfalls (full detail in reference):
- #11 Dispatch-fail → parent-direct fallback (don't retry
delegate_task blindly)
- #13 Silent truncation of depp-models — use sentinel+structure validation
- #16 Queen-Lane-Plan for structured multi-lane missions (5.4x speedup)
- #17 Parent-direct for <30min lanes — subagent context-burn too costly
- #18 Phase-0-Inventur FIRST — grep what's already in repo
- #20 Verify user-path before
ls — user-path is a hint, not a contract
Cross-Reference: hermes-orchestration (V2.4+)
Seit 2026-06-27 existiert hermes-orchestration als dedizierter Ausführungsskill — dieser Skill bleibt das Pattern-Repository, der Ausführungsskill ist hermes-orchestration.
| Feature | multi-agent-orchestration (Pattern) | hermes-orchestration (Runtime) |
|---|
| 3-Expert Deep Research | ✅ Phase 1 Pattern | ✅ Task-Typ research |
| Queen-Bee Delegation | ✅ Konfig + Owl Alpha | ✅ Hermes Subagent Bridge |
| Parent-Direct-Fallback | ✅ Pitfall #11 | ✅ Automatisch + --manual |
| Heuristic Extraction / Mnemosyne / Cron | ❌ | ✅ |
Empfehlung: Für neue Runs hermes-orchestration nutzen (es läuft). Diesen Skill für das Pattern-Wissen und die documentierten Pitfalls laden.
Hermes Subagent Bridge
For the canonical code pattern (Worker-Role → Toolset mapping, dispatch with parent-direct fallback, worker-contract parsing): see references/hermes-subagent-bridge.md. Used by hermes-orchestration (V2.1) and the-dmz-transfer.
See Also: Related Skills
hermes-orchestration (same category) — the runtime with orchestrator.py, metrics_extractor, heuristic_aggregator, mnemosyne_importer, weekly-pipeline cron. Use this for actually running orchestration.multi-agent-pitfalls-cheatsheet — load BEFORE every delegate_task for the top-20 trigger-watchlistsubagent-driven-development — execution pattern using subagentsskill-library-maintenance — class-level skill for slim-down patterns and 4-Tier Verification
Documentation Planning — 4-Agent + Queen Pattern
For creating/improving README, DESCRIPTION, NAVIGATION docs across a directory tree. Spawn 4 parallel sub-agents (README, Description, Navigation, Control) + Queen consolidation. Full role definitions, context templates, toolset assignments, risk categories, verification scripts: see references/documentation-planning-pattern.md.
Workflow
- Load
multi-agent-orchestration (this skill) for the pattern
- Load
multi-agent-pitfalls-cheatsheet BEFORE every delegate_task for traps
- Spawn 3 experts per Phase 1
- Cross-check scout claims per Phase 3.5 (verify, don't trust)
- Synthesize + present 2-4 verified fixes per Phase 3.6 (user decides)
- Write final report per
report-template.md
Live demos & known-good baselines: see multi-agent-pitfalls-cheatsheet references (demo-session-2026-07-02.md, 2026-07-02-additions.md).
Converted from Hermes-Skill by Yuno. Original SKILL.md preserved in references/original-hermes.md.
