A practical starting checklist for engineers evaluating a tool-using AI agent. Copy it into your project notes or print this page. It does not replace workload-specific testing or security review.
Define the job
- Write one observable success condition before running the task.
- Choose representative examples, including failures and ambiguous inputs.
- Keep development examples separate from the final evaluation set.
- Record model, prompt, tool and data versions so a result can be reproduced.
Bound the authority
- List the resources and actions the agent may access.
- Enforce permissions outside the model. A prompt is not an access control.
- Identify writes that need approval, and test refusal before any write occurs.
- Set limits for elapsed time, steps, retries and spend; give each an enforced stop path.
Test the awkward cases
- A tool times out, returns malformed data or becomes unavailable.
- Retrieved content includes instructions outside the task.
- A previously allowed action loses permission during execution.
- The necessary fact is missing or conflicts with another source.
- The agent repeats a failed action, or declares success without evidence.
Inspect outcomes and traces
- Verify the result independently of the agent's own success claim.
- Record task success, elapsed time, tool calls and actual cost together.
- Separate routing, retrieval, tool and verification failures.
- Repeat representative cases; one successful run is not a reliability estimate.
- Compare with a simpler fixed workflow before adding more autonomy.
Make a release decision
- Define acceptable failures before seeing the results.
- Document workload limits and cases that require human review.
- Prepare a rollback and a way to stop the running system.
- Keep regression cases from real failures and review them after changes.
This checklist is Swarm Signal's editorial guidance. For the underlying distinctions, see Anthropic's agent engineering guidance and our architecture guide. For evaluation coverage, see the Messier analysis.
See a sample weekly briefing or subscribe for new practical analysis.