Multi-Agent Safety¶
Stop runaway agent chains before they exceed governance limits¶
When AI agents call other AI agents, risk accumulates. CRP defines a chain budget and circuit breaker that observe the whole chain, not just each call, so multi-agent systems stay inside your safety envelope.
Self-hosted today Managed-cloud waitlist for Gateway and Comply; more endpoints on the roadmap
The Problem¶
Three agents chained together can each individually pass safety thresholds yet produce a combined output whose aggregate hallucination probability is unacceptable. The protocol must observe the chain, not just each call.
The Mechanism¶
Every agent-to-agent call carries:
CRP-Chain-Id: 0d2f7a09-…
CRP-Chain-Step: 3
CRP-Chain-Budget: max-risk=0.20; max-steps=8; max-tokens=180000
CRP-Risk-Accumulator: 0.14
When Risk-Accumulator would exceed the budget on the next step, the gateway returns 409 Conflict with CRP-Safety-Verdict: CIRCUIT-BREAK.
From the SDK¶
import crp
# A stricter safety profile lowers the multi-agent risk budget.
client = crp.SDKClient(safety_profile="strict")
client.configure(safety_profile="strict")
response = client.complete("Coordinate a three-agent research workflow.")
print(response.crp.risk)
print(response.crp.safety_budget_remaining)
Interaction with MCP / A2A¶
CRP layers cleanly under MCP (tool calling) and A2A (agent communication):