the agent debugging process (small group edition)
how this works
Everyone brings one real example of an agent doing something wrong or unexpected. Your group will take turns: during your turn, the whole group works together to figure out why your agent failed. One timer runs for each person's turn. When the timer ends, the group moves to the next person.
The group's job is diagnosis: finding the most likely cause. At the end of the exercise, everyone gets time on their own to try fixing their problem based on what the group found.
pick your problem (3 minutes, on your own)
Pick one time this week when your agent did something wrong or unexpected. Open the actual conversation where it happened. Write your problem in one sentence next to your name in the setup section of your group's tab in the worksheet — what you asked for, what you got.
If you cannot find the conversation, ask the agent to help you find it. If you still cannot find it, that is a real problem too — bring that to your group.
If nothing went wrong for you this week, bring the workflow or automation you trust the least.
group diagnosis (5 minutes per person)
Reset and restart the timer for each person, so that everyone gets a turn.
The person whose turn it is shares their screen. The others ask the questions below and make sure the person actually looks at each thing on screen, instead of answering from memory. Go through the steps in order — the worksheet has a diagnosis block for each person that follows these same steps; someone other than the person on the hot seat fills it in as you go. Start the timer at the beginning of each turn, and switch to the next person when it ends.
1. look at what happened
Show the exact conversation: what you asked for, and what the agent did instead. If you only have a general feeling that the agent "is not working well," stop and find one specific example before going further.
2. read what the agent said
Read through the agent's messages in that conversation together. The explanation for the failure is often written right there. Then ask the agent directly, while the group watches: "Explain in simple terms what you did here, what information you could see, and why you did it this way."
3. check the five places problems live
The instructions. Read the actual prompt, automation, or instructions the agent was given, word for word. Do they point to the right files and folders today? Is it clear what a finished result looks like?
The skills. If the agent was supposed to use a skill, does that skill actually exist? Is it saved in the folder where this tool looks for skills? Was the conversation started in a place where the agent can see it?
The files and data. Does the agent's computer actually have the file it needed? Are there multiple copies of the file in different places, and if so, which one is the real one? If files are supposed to sync between places, is that sync actually working right now?
The actions. Is the agent actually able to do the thing you asked — for example, edit that type of file, or reach that account? Is it connected and logged in? Watch out for the agent pretending to succeed, like creating a new file instead of editing the one you asked about.
The platform. Sometimes the agent product itself is buggy, and nothing in your setup is wrong. The signature: your instructions, skills, files, and connections all check out, the error message looks opaque or internal (a raw code error, nothing useful in the logs), and running the same thing again works. The fix is different too — report it to the vendor, work around it, or wait, rather than changing your setup.
4. agree on the most likely cause
End the turn with one sentence the whole group agrees on: "It most likely failed because ___." Write it in the worksheet and read it aloud — the person whose agent it is must agree it is the most likely cause before the group moves on. There can be more than one cause — if you found one, ask whether there might be a second one underneath it.
try the fix (5 minutes, on your own)
Everyone works alone on their own problem. Take the cause your group found and try to fix it at the source: change the instructions, move the file, fix the sync, connect the account. If it turned out to be a platform bug, report it to the vendor and note your workaround. Do not just tell the agent to do better in the chat — a correction typed into one conversation will not change what happens next time. Fix the thing the next run will actually read.
Then check whether the fix worked, if you can: run the task again, or try a small version of it. Record in the worksheet's troubleshoot section what you changed and how you verified it.
If you get stuck, ask your group first, then the instructor.
share with the class (5 minutes)
Which groups found the cause of a problem? Which people got unblocked? A few people share, thirty seconds each: what the problem was, where the cause turned out to live (instructions, skills, files, actions, or the platform), and what the fix was. If you can, re-run your original request live and show what happens now — that is the finish line.
Before leaving, fill in the worksheet's share-back section for your own problem: where the problem lived, what fixed it, and how you measured success.