Multi-Agent Composition (CALIBRE-Compose)
Graph-based scoring of multi-agent AI systems. Complements the per-agent CALIBRE.
What it is
Paper Section 14 (forthcoming). CALIBRE-Compose scores the composition graph of multiple agents as a whole. Seven compositional risk classes CC-01 through CC-07 catch problems that no individual-agent assessment can see: cascade failure, trust chaining, authority escalation, feedback loops, context contamination, handoff injection, emergent coordination.
When to use
- Any deployment involving two or more AI agents talking to each other.
- LangGraph, AutoGen, CrewAI, MCP-connected agent networks.
- Enterprise LLM agents integrated with SaaS agents (Salesforce, SAP, ServiceNow).
Walk-through
- 1Sidebar: Regulatory > Multi-agent systems.
- 2Click '+ New composition'.
- 3Describe nodes (agents, MCP tools, shared memory, humans, external systems) and edges (invokes, reads, writes, handoff, trusts, verifies) as JSON.
- 4Set context flags: iterative review, untrusted external inputs, persistent memory.
- 5Score. If C-GII >= 40, the linked initiative's Compositional-Coupling amplifier auto-fires.
Common pitfalls
- Trying to model hidden or undeclared agents in the current version. Add them as external nodes with trust_level=untrusted; still incomplete coverage (see scenario in Help > Scenarios > 'Hidden and unlabelled agents').
- Treating all edges as 'trusts'. Add 'verifies' edges to represent where actual verification happens, otherwise trust-chaining score is inflated.