---

name: firm-orchestration version: 1.0.0 description: > Pyramid multi-agent orchestration for OpenClaw: routes objectives from a CEO agent down through departments, services and employees via sessions_send / sessions_spawn, collects and merges results, enforces handoff contracts, and writes the final deliverable back to the originating session. author: romainsantoli-web license: MIT metadata: openclaw: requires: env: [] bins: [] tools: - sessions_list - sessions_send - sessions_spawn - sessions_history primaryEnv: "" tags:

• orchestration
• multi-agent
• a2a
• firm
• pyramid
• enterprise

---

firm-orchestration

This skill implements the A2A (Agent-to-Agent) pyramid pattern for OpenClaw.

Architecture

CEO Agent (orchestrator)
 ├── Department Strategy
 │   └── Service Planning → Employee Analyst
 ├── Department Engineering
 │   └── Service Backend  → Employee Implementer
 ├── Department Quality
 │   └── Service Testing  → Employee Auditor
 └── Department Operations
     └── Service Release  → Employee Coordinator

Usage

Send this to your OpenClaw session to trigger a full firm orchestration run:

@firm-orchestration run
  objective: "Build a payment API"
  departments: ["engineering", "quality"]
  delivery_format: "github_pr"

Tools activated

Tool	Purpose
sessions_list	Discover active department/service sessions
sessions_spawn	Spawn missing sessions per pyramid level
sessions_send	Delegate objectives down the hierarchy
sessions_history	Collect results from child sessions

Handoff contract

Each delegation payload follows this schema:

{
  "from": "ceo",
  "to": "department:engineering",
  "objective": "...",
  "constraints": ["...", "..."],
  "definition_of_done": "...",
  "context_ref": "memory:delivery/latest",
  "reply_session": "main"
}

Merge strategy

Results from all departments are:

1. Collected via sessions_history with a 30-second deadline
2. Deduplicated by objective_key
3. Merged in dependency order (Strategy → Engineering → Quality → Ops)
4. Formatted according to delivery_format

Operating Protocol (Anthropic-style)

Based on real Anthropic team practices — "How Anthropic teams use Claude Code"

Phase 1 — Parallel dispatch (never sequential)

Fan-out simultaneously to all departments via sessions_send. Never wait for one department before launching the next. Each session receives the full handoff contract and maintains its own complete context. Store all reply_session refs for convergence.

Objective received →
  sessions_send(engineering) ‖ sessions_send(quality) ‖ sessions_send(ops) ‖ sessions_send(strategy)
→ wait(deadline=30s)
→ collect via sessions_history

Phase 2 — Iterative loop on blockers

If a department returns status: blocked, do NOT resolve it yourself. Spawn a joint resolution session with the two conflicting departments and let them iterate:

engineering blocked by legal →
  sessions_spawn(participants=[engineering, legal], objective="resolve_blocker") →
  wait(max_iterations=2) →
  collect resolution

Maximum 2 re-delegation cycles before escalating to CEO with explicit blocker report.

Phase 3 — Convergence with partial acceptance

30-second hard deadline. After deadline: accept partial results, mark missing department outputs as status: timeout, include them in final report as open items. Never block delivery on a single department.

Phase 4 — Validate before merge

Before merging each department output into the final deliverable:

1. Check output satisfies its definition_of_done
2. If DoD not met: flag as quality: partial — do not silently drop
3. Merge in dependency order only: Strategy → Engineering → Quality → Ops

Phase 5 — Deliver + document

After every completed orchestration, automatically append:

1. Run summary (1 paragraph)
2. Departments that delivered / timed out / were blocked
3. Architecture/process decisions made
4. Suggestions for improving the next similar run

All final outputs carry the mandatory disclaimer:

⚠️ Contenu généré par IA — validation humaine requise avant utilisation en production.

Phase 6 — Git checkpoints (when Engineering is involved)

Require Engineering to commit after each sub-task — not only at end of run. Reject PRs that are not draft + labelled needs-review. Never allow direct merge to main.

Security

• All inter-session calls use reply_session: "main" to avoid orphaned sessions
• sessions_spawn is rate-limited: max 20 spawns per orchestration run
• Payloads are validated against the handoff schema before dispatch
• No external network calls — pure Gateway WebSocket routing

Example prompt

Use the firm-orchestration skill to:
  objective: "Audit the authentication module"
  departments: ["quality", "engineering"]
  constraints: ["read-only access only", "no production changes"]
  definition_of_done: "Security report with CVSS scores and fix recommendations"
  delivery_format: "markdown_report"

---

💎 Support

Si ce skill vous est utile, vous pouvez soutenir le développement :

Dogecoin : DQBggqFNWsRNTPb6kkiwppnMo1Hm8edfWq