Trace autopsy
An AI agent's run went wrong. Several steps look broken — but only one is the root cause; the rest are just symptoms that inherited the bad state. Click the step where it first went wrong, then see the autopsy: the root cause and every consequence it produced. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the agent's trace step by step — each action and its output.
- Click the step where reality and the plan first diverged (not the last one that looks wrong).
- Submit and read the autopsy: which step was the root cause, and how many later steps were just symptoms.
What this clears up (the fundamentals)
- The last error isn't the cause — an agent throws where the bad state finally becomes fatal, often several steps after the real mistake.
- One bad decision cascades — a wrong tool argument or misread early on poisons everything downstream.
- Debug backward — walk the trace from the error back to the first divergence; that's the step to fix.
- Symptoms mislead — patching the step that threw leaves the root cause in place, so it breaks again.
Where it's used
An AI-engineering debugging drill — the fastest way to build the habit of reading agent traces for root cause instead of last symptom. It's a miatz build-lab concept playable; learn to build it yourself.
FAQ
Why isn't the failing step the problem?
Because it's usually the point where the bad state finally becomes unrecoverable, not where it started. The mistake happened earlier and flowed downstream until something choked on it.
How do you find the root cause?
Walk the trace backward from the error to the first step whose output doesn't match what a correct plan would produce. That first divergence is the cause; everything after is a consequence.
Why do the downstream steps look wrong too?
Because they operated on the corrupted output of the root-cause step. They did their job correctly on bad input, so their results are wrong through no fault of their own.
Is anything uploaded?
No. The trace and diagnosis run entirely in your browser — nothing is transmitted or stored. Turn off your Wi-Fi and it still works.
Limits
A teaching model with a fixed seeded trace and one root error — real agent runs have branching and multiple faults. The root-cause reflex it trains transfers directly.
Related
Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.