What verdict and next step should an AI SRE record for error spikes?

Tares lets an AI SRE end an error-spike run with the conclude tool. When one entity looks like a problem, record the verdict investigate, a headline of up to 100 characters and a next step of up to 300 characters. When there is nothing to report, record no_op, which creates no finding. The verdict words are yours, so you define them on the agent.

How it works

An agent set to end with a conclusion gets the conclude tool, which ends a run on purpose. The tool has two outcomes:

• no_op: there is nothing to report, and no finding is recorded.
• A finding: a verdict about a named entity, plus a headline and a next step.

The two extra fields of a finding work like this:

• headline: one line of up to 100 characters. It says what was found in the reader’s terms.
• next_step: up to 300 characters. It says what a person should do, and it is empty when nothing needs doing.

The project Overview shows the headline and next step on each card, and agents are told to give both. If a run does not provide them, Tares takes the headline from the note’s heading, bold lead or first sentence. It takes the next step from the text after “Next step:”, “Next:” or “Recommendation:”.

The verdict words come from you. Each is one word (letters, digits, dashes), and you can list up to 10 on an agent. With a list, the tool accepts only those words and tells the model what each means. A finding must carry one, and an off-list word or a missing verdict goes back to the agent to fix within the same run. Without a list, a finding can carry any verdict, or none. See How a run ends for the full rules.

Set it up

Set the verdicts on the agent that watches your error signals. This follows the AI SRE flow described in AI SRE.

1. Open the agent form and go to the How it ends section.
2. Tick Ends every run with a conclusion. In YAML this is concludes: true. The agent then always gets the conclude tool, and Tares adds the instructions for it, so your prompt only describes the job.
3. Under Verdicts it can give, add investigate with a short note on when to give it. For error spikes, that is when one entity looks like a problem and needs a closer look.
4. In your prompt, tell the agent to name the entity that is spiking in the headline and to put the action a person should take in the next step.
5. If a second agent takes over after a handoff, give it rca for when you found what is failing and why, and resolved for when it was noise or is already over.

The triage agent from the docs looks like this:

agents:
  - name: triage
    trigger: watch-web
    concludes: true
    verdicts:
      - { verdict: investigate, when: one entity looks like a problem and needs a closer look }
    prompt: |
      ...

The Triage (on a schedule) preset already comes with investigate. The Root cause after a handoff preset comes with rca and resolved. Your agent’s handoffs key off these verdict words, so keep them consistent between agents.

If a run stops without concluding, Tares asks once more with conclude as the only choice. That way a run on an error spike does not end without a verdict.

What you get

• A clear verdict per run. An error spike on a named entity ends as investigate, rca or resolved, or as no_op when there is nothing to report.
• Cards a person can read at a glance. The headline says what was found in one line, and the next step says what to do. When nothing needs doing, the next step stays empty.
• No finding for an empty run. no_op records nothing, so a run with nothing to report adds no card to the project.
• Enforced vocabulary. With a verdict list, the agent cannot invent words or skip the verdict, because the tool sends a wrong or missing one back within the same run.
• Handoffs that work. Because verdicts are fixed words, a later agent can pick up an investigate finding and finish with rca or resolved.

When something else fits

• If you do not tick the conclusion option, the agent’s last message is its finding and carries no verdict. That suits a simple summary where nothing downstream reads a verdict.
• If you leave the verdict list empty, a finding can carry any verdict or none. Use that for exploration, then add a list once the words settle.
• If a spike needs no action, give an empty next step instead of inventing one. If there is nothing to report at all, end with no_op.

Related