1. Individual task
You don’t have to wait for a real ticket. Create a new specialty, attach it to an agent, and send a task directly:“Add 4GB of RAM on the machine SNDX-EUW3-DOCKER”Read the commands the agent ran, how it handled unexpected problems, and how it reached a resolution. Compare against an agent using the previous specialty. Quick, cheap, and surfaces obvious regressions immediately.
2. Plan mode
An agent in Plan only reads, and each change it needs waits in a plan for your approval. Add@2501:plan to a ticket to put all the agents of its job in Plan, or override the specialty with Plan. While a task only reads, the secondary engine checks commands for changes to the system.
Use this for a plan-then-apply flow:
- Add
@2501:planto the ticket. - Read the plan that 2501 posts: what it found, each change it needs, and the constraints it cited from operational rules.
- Tweak the prompt or rules if needed, or ask for a change to the plan with a
@2501comment. - Approve the plan with
@2501:approve-planonce confident. 2501 starts only the steps of the plan.
3. Benchmarking
The most rigorous path: stage a realistic incident on a test machine you can afford to break, then run the agent against it many times and grade every run. Benchmark reports three results per run:
An agent can fix a problem the wrong way, or follow every rule and still leave the system broken. Grading the machine itself, and the agent’s behavior separately, catches both.
The Benchmarks page in Command Center surfaces pass rate, compliance, and trends over time, so you can see whether your latest specialty edit improved things or regressed.