All learning paths Learn See Try Make Check the steps, not the ending What this lesson is about A right answer reached by a wrong step will be wrong next time. Mark the step, not the result.
An agent's train-fare run ends on the figure the railway site shows, but step three opened a 2024 table. What do you file?
A Pass the run and add a note that step three used an old table. B Check the final figure against the railway site again, and pass it if the two agree. C A right answer reached through a wrong step is a failure, not a pass.
Check my recall Why this matters A right answer reached through a wrong step is luck, and luck does not repeat across a queue of runs. The Hugging Face course teaches watching a run by its own steps and error traces. That is what turns one lucky pass into a bug somebody can fix.
Judge a run by its steps, and file the first bad one instead of the final answer.
A right answer reached through a wrong step is luck, and luck does not repeat across a queue of runs. The Hugging Face course teaches watching a run by its own steps and error traces. That is what turns one lucky pass into a bug somebody can fix.
A right answer reached through a wrong step is a failure, not a pass. Check each step against what it claimed to do, not against the final answer. Mark the first bad step; everything after it is a consequence of that one. Task Check one agent run about train fares where the final answer is correct.
Weak approach Opens the railway site and checks today's fare against the agent's answer. They are equal, so the run is marked a pass. The six steps behind that answer are never opened.
Stronger approach Files step three as a break: it opened a fare list from last year instead of this year's. Adds that the matching answer was luck and will not repeat.
Why the stronger approach works The weak check judged the ending. The strong one judged each step against what it said it was doing.
Try a changed situation You are given only an agent's step log. The final answer has been removed. Can you still judge the run?
Make something yourself Take one agent run and review it step by step. Name the first bad step, what it actually did, and the step that should have been there instead.
Lesson 4 of 9 on Running and checking AI agents next Decide where a run must stop Part 2 ends in a work sample you can send
Read the primary or official source: Hugging Face Agents course · watching and judging agents