Agent errors
Funny once. Less funny by the fourth retry. Eight fictional examples with checks you can apply to your own workflows.
The log avalanche
Tools
You asked for one error. I brought the director’s cut.
Symptom. A full export is read repeatedly for one question, with no filter or saved observation.
What to check. Which slice answers the question, is it complete, and where is the evidence?
Open the technique · Russian →The amnesiac handoff
Context
Everything’s in the notes. Except why we’re here.
Symptom. The next session gets a recap of actions, but loses the goal, constraints and unfinished verification.
What to check. Can the next operator continue without guessing or repeating the investigation?
Open the technique →A medal for a cache hit
Caching
The hit rate looks lovely. I haven't opened the bill.
Symptom. A cache hit rate stands in for whole-task savings without accounting for the other calls or outcome quality.
What to check. What did the source measure, and which costs did it leave out?
Open the technique · Russian →One more thing, forever
Retries
All done. Give me another minute.
Symptom. The task has no completion criterion and new attempts have no testable goal.
What to check. Has acceptance passed, and what would another attempt change?
Open the technique →The tool buffet
Tools
I asked for a screwdriver. They brought the garage.
Symptom. A small task loads definitions for a large tool library.
What to check. Which tools does the agent need now, and can it find the others later?
Open the technique · Russian →The retry carousel
Retries
Perhaps the same error will have a change of heart.
Symptom. The agent repeats a failed action without changing the inputs or diagnosing the failure.
What to check. What changed after the last attempt, and why should this one work?
Open the technique · Russian →The busywork parade
Parallel work
Everyone’s busy. Has anyone seen the result?
Symptom. Several branches run, but their contribution to the usable result is unknown.
What to check. Which results did you use, what did the discarded branches cost, and where was the critical path?
Open the technique · Russian →The short-answer trophy
Evaluation
Fewer tokens. The right answer is in the next release.
Symptom. Lower usage is treated as success before the output passes a quality check.
What to check. Did it meet the same acceptance criteria, with every attempt counted?
Open the technique →