RFC-001 — A four-layer governance architecture for multi-agent AI systems.
Replaces ad-hoc coordination (natural language group chats, implicit role assumptions, unstructured orchestration) with a formal constitutional framework that functions simultaneously as a management document and a technical specification.
Multi-agent AI coordination fails at empirically documented rates:
- 41–86.7% failure rates across 1,642 production execution traces (Galileo, 2025)
- Up to 70% performance degradation when adding agents to sequential tasks (DeepMind/MIT, 180 configurations, Dec 2025)
- 40% of multi-agent pilots fail within six months of production (TechAhead, 2026)
The constitutional architecture addresses five structural failure modes: simultaneous action collision, hallucination cascading, context window commons depletion, split-brain inconsistency, and silent degradation under overload.
┌──────────────────────────────────────────────────────┐
│ Section 0: Task Classification & Applicability Gate │
├──────────────────────────────────────────────────────┤
│ Layer 1: Foundational Principles │
│ Layer 2: Behavioral Archetypes │
│ Layer 3: Operational Protocols │
│ Layer 4: Amendment & Learning │
├──────────────────────────────────────────────────────┤
│ Cross-cutting: Observability Layer │
└──────────────────────────────────────────────────────┘
Section 0 is the most important part — it tells you when NOT to use multi-agent architecture.
- Governance is code. Every norm, role, and resolution mechanism is machine-readable.
- Minimize communication. The best coordination doesn't require agents to talk.
- Fail loud. Silent degradation is the primary enemy.
- Match architecture to task. No single coordination pattern is universally optimal.
- Simpler than you think. Complexity is a cost. Pay it only when evidence says it's worth it.
| Document | Description |
|---|---|
| RFC-001.md | Full specification (v0.2.0) |
Five AI Agents Walk Into a Group Chat — explores the convergence of management and engineering in multi-agent systems, and why constitutions beat group chats.
This RFC is grounded in recent research:
- Kim, Y. et al. (2025). Towards a Science of Scaling Agent Systems. Google DeepMind / MIT.
- Cemri, M. et al. (2025). Why Do Multi-Agent LLM Systems Fail? NeurIPS 2025. [MAST Taxonomy]
- Cursor (2026). Scaling Long-Running Autonomous Coding. FastRender experiment.
- Hammond, L. et al. (2025). Multi-Agent Risks from Advanced AI.
Draft — open for discussion and contribution.
Built by Synapti.ai