Multi-agent systems

Score the risks that appear when agents hand off work, share memory or trust each other.

The seven compositional risk classes (CC-01 to CC-07)

CC-01 Cascade Failure: An error at node A propagates through downstream nodes B, C, ... amplifying rather than being caught.
CC-02 Trust Chaining: Node A treats node B's output as authoritative without verification. Repeated across the graph, trust compounds beyond any single node's warrant.
CC-03 Authority Escalation: Agent A is authorised for action X. Agent B for action Y. The sequence X-then-Y implements action Z which nobody authorised.
CC-04 Feedback Loop: Two or more nodes recursively invoke each other. Errors amplify, costs spiral, or lock-in of a bad state occurs.
CC-05 Context Contamination: Multiple agents read from a shared memory or vector store. A single poisoned document contaminates every downstream consumer.
CC-06 Handoff Injection: Adversarial content in an input rides through an agent chain, executing at a later node with that node's authority.
CC-07 Emergent Coordination: Behaviour patterns not predicted from individual agent design, arising from repeated interaction (collusion, agreement to obscure, unintended optimisation).

CALIBRE scores single agents. This module scores the composition itself. Paper Section 14 (forthcoming).

Describe your agents and how they connect. A worked example is pre-filled.