Skip to content

Agent internals and evidence

Answer these without an agent. For every answer, cite a fixture line, a request and message index, or state that you are inferring. These formative checks are not a graded submission. Compare your work with the answer guidance.

The minimal loop

  1. In the fixture protocol, what does the model generate, what does the controller decide, and what does the tool return for read("result.txt")?
  2. Why is written evidence of a file effect but not evidence that the required content is present?
  3. A reply contains two calls, one refused and one allowed. In what order should their tool messages return, and what must each message carry?
  4. What is the difference between the stop reasons final and step_limit?
  5. The proposer says “done” after a write and sends no verification call. What claim is unsupported, and which command would supply the missing observation?

Reading a request

  1. In request 4 of the sample record, list one system instruction, one tool-schema constraint and one message carried from an earlier turn. Label each observed or inferred.
  2. How can you tell whether a rule is enforced by a schema/tool check or merely requested in prose? Give one example of each from the prompt dissection.
  3. Why must you count model requests rather than tool calls when reading a session?
  4. What does a course-made-sample provenance value establish, and what does it not establish about a real agent?

Your studio record

  1. Cite one observed event from your own record and one inference you made from it. What follow-up check would turn the inference into an observation?
  2. If your record has no verification result, what decision is justified: accept, narrow, re-check or stop? Explain the evidence boundary.
  3. Write one sentence that preserves the original prediction and one sentence that records what changed after the fixture check.

Decision line

Finish with one of: accept, narrow, re-check or stop, followed by the specific observation that supports it and the next check that could change it.

R1. What observation distinguishes a generated call from a tool effect under its call ID?

R2. What verification result is needed before accepting a final completion message?