CrewAI
Governance
CrewAI orchestrates multi-agent teams where agents delegate, collaborate, and act autonomously. More agents means more risk. OnLeash governs every crew member individually and the crew collectively — per-agent policies, delegation governance, and crew-level kill switches.
Multi-Agent Risks
CrewAI's power comes from multi-agent orchestration. That same orchestration creates compound risks.
Inter-agent escalation
One crew member delegates to another, which delegates further. Privilege and scope expand with each hop. Without governance, a low-risk initial task can cascade into high-risk operations.
Uncoordinated tool access
Multiple crew members accessing the same tools simultaneously. Race conditions, duplicate writes, conflicting operations — none visible until damage is done.
Cost amplification
Hierarchical crews multiply LLM calls. A crew of 5 agents running 3 tasks each can execute 50+ LLM calls per workflow. Without metering, costs spiral.
Accountability gaps
When a multi-agent workflow fails, which agent caused it? CrewAI logs task completion but does not provide per-agent action-level audit trails with tamper-proof integrity.
Blast radius
A compromised crew member can influence the entire crew through shared context and task delegation. Without containment, one injection can propagate across all agents.
Governance for Multi-Agent Systems
Per-agent policies
Each crew member gets individual governance policies. The researcher agent has different permissions than the writer agent. Least privilege, enforced.
Delegation governance
When one crew member delegates to another, the delegation event is governed. Policies can restrict delegation chains, require approval for cross-role delegation, or block delegation entirely.
Shared tool governance
Tools shared across crew members are governed at the tool level. Concurrent access policies prevent race conditions. Per-agent tool restrictions enforce least privilege.
Crew-level kill switch
One kill switch halts the entire crew. Per-agent kill switches halt individual crew members. Circuit breakers auto-trigger on crew-level anomalies.
Cross-agent audit trail
Unified cryptographically chained audit log across all crew members. Trace any action back to the originating agent, task, and delegation chain. Tamper-proof.
Cost controls
Token metering per agent and per crew. Budget caps with automatic circuit breakers. Real-time cost visibility in the dashboard.
How It Works
The OnLeash Python SDK provides per-agent callback handlers for CrewAI. Each crew member gets its own governance identity and policies. The SDK plugs into CrewAI's callback system — no changes to your crew definition or task logic.
- 1. Install the Python SDK
- 2. Add a governance callback to each crew member
- 3. Each agent gets individual policies — different roles, different permissions
- 4. Every tool call by every agent is governed before execution
Individual policies per crew member. Unified audit trail across the crew.
Govern Your Multi-Agent Crews
Free tier. Python SDK integration. Every crew member governed individually.
Deploy Crew Governance