All learning paths Learn See Try Make Watch one run from end to end What this lesson is about Sit through a whole run without touching it, and write down every step in your own words as it happens.
An agent at your office begins a billing run you have never watched, and nobody else is free. What do you do?
A Watch the first few turns, then step in and correct it before it wastes the hour. B Watch one whole run without stepping in, and write down each step as it happens. C Let it finish and then read its own log from the top, which holds every step.
Check my recall Why this matters You cannot check a run you have never watched, and a run you stepped into halfway is a different run. Watching one from end to end turns an agent from a box into a sequence you can argue with. A trace is the record the agent's tool prints, one line per step. One step record, in the shape you will meet it. Step 3. tool: search. args: file "March rates". result: empty, 0 rows. status: success. 840 ms. tokens: 1,204 in, 12 out. Two things a trace shows that a summary never does. A step whose result was empty while its status said success, like the one above. And a retry loop: the same call at steps 4 and 5, with the same arguments, seconds apart.
Sit through one whole run and write a record of it that somebody else can act on.
You cannot check a run you have never watched, and a run you stepped into halfway is a different run. Watching one from end to end turns an agent from a box into a sequence you can argue with. A trace is the record the agent's tool prints, one line per step. One step record, in the shape you will meet it. Step 3. tool: search. args: file "March rates". result: empty, 0 rows. status: success. 840 ms. tokens: 1,204 in, 12 out. Two things a trace shows that a summary never does. A step whose result was empty while its status said success, like the one above. And a retry loop: the same call at steps 4 and 5, with the same arguments, seconds apart.
Watch one whole run without stepping in, and write down each step as it happens. A log hides two things: how long each step took, and where it went back and retried. Say afterwards which single step you would have done differently, and why. Task Watch one agent run on sample invoices and write it up for the shop owner.
Weak approach Watches the whole run and writes that the agent filled the invoices. An old rate file was used somewhere in the middle. No step is numbered and no time is written down.
Stronger approach Copies the trace's six steps and rewrites each in plain words with its time. Step 3: search for the March rates file, result empty, status success, 840 ms. The old file stayed in use. Steps 4 and 5: the same save call twice, same arguments, 2.1 s and 2.3 s. Names step 3 as the one to change, because an empty result was treated as done.
Why the stronger approach works The weak note cannot be acted on. The strong one names the step, its empty result, the time and the retry, so the fix is obvious.
Try a changed situation A colleague hands you her written report of a run. Every step is listed and it all looks fine. What do you ask for?
Make something yourself Watch one complete run without stepping in. Write every step in your own words with the time it took. Mark any retry, and name one step you would change.
Lesson 3 of 9 on Running and checking AI agents next Check the steps, not the ending Part 1 ends in a work sample you can send The rules this work sample is read against
Read the primary or official source: Hugging Face Agents course · the agent loop