Checklist
Agent Safety & HITL Checklist
Controls checklist for tool-using agents and multi-agent workflows in enterprise environments.
Use this in workshops, security reviews, and delivery planning. Print or copy freely.
Tooling
- ✓Every tool has an allowlisted purpose and scope
- ✓Write actions are separated from read exploration
- ✓Credentials are least privilege and rotatable
Human control
- ✓High-risk actions require human approval
- ✓Reviewers see enough context to decide quickly
- ✓Timeouts and escalations are defined
Observability
- ✓Tool calls and decisions are logged
- ✓Cost per task is visible
- ✓Stuck or looping workflows are detectable
Release readiness
- ✓Task success metrics exist
- ✓Unsafe action tests are in the evaluation suite
- ✓On-call / owner model is clear after handoff