Scoring an agent on the AI Agents track. Score tasks, not vibes. A task has a start state, a hidden correct outcome, and a grader. The grader checks the world, such as the database row, not whether the reply sounded confident.
This lesson assumes you already worked through Boundaries around actions.
The idea in practice
Build a set of tasks with fixtures. Run the agent. Compare final state to the expected state. Also record step count, tool errors, and whether a person was asked.
A concrete check
goal = {
'track': 'AI Agents',
'lesson': 'Scoring an agent',
}
checks = [
'input available at decision time',
'score matches the real decision',
'failure case written down',
]
print(goal['lesson'])
for item in checks:
print('-', item)
Run the sketch locally if you have Python. The printout is a reminder of the checks, not a trained model. Replace the strings with the real inputs from your own example before you treat it as a design.
What usually goes wrong
Grading only the final sentence misses a wrong tool call that the prose then covers up. When this happens, stop adding parameters or tools. Fix the check, the data, or the permission, then run the same example again.
What to write down
- The input you are allowed to use at decision time.
- The output and the score or pass rule.
- One failure you will test on purpose.
- What you will not claim the system can do.
Practice
Write two tasks with a fixture and a pass rule that looks at data, not wording.
Self-check
- Say Scoring an agent in one sentence that mentions an input and an output.
- Name the failure mode in this lesson and the check that would catch it.
Done when: you can explain this lesson without the page open, and you have a written failure case.