AGENT_AUTOPSY//
// CASE FILES 01–17 · RUNNABLE · TESTED

THE AGENT
AUTOPSY

17 ways AI agents fail in production. Each one runnable. Each one fixed.

Most agent tutorials show you how to build one. This series shows you what breaks after you do.

START HERE → FIELD CHECKLIST
87 tests 17 folders 0 API keys required By Lawrence Jefferson II (Menoko OG)
TTY0 · 01-runaway-loop
$ python broken.py
IDLE

Mock model. Output matches the repo's broken.py / fixed.py.

// INDEX

The 17 failures

L/C
CATEGORY
Loop and control
FAILURE 01 /01-runaway-loop The runaway loop The agent is the only thing deciding when to stop. broken.pyfixed.pytests FAILURE 02 /02-stuck-agent The stuck agent The work is done and the done-check can't see it. broken.pyfixed.pytests FAILURE 03 /03-confidence-collapse Confidence collapse "are you sure?" on repeat flips a right answer. broken.pyfixed.pytests
T/I
CATEGORY
Tools and integration
FAILURE 04 /04-tool-hallucination Tool hallucination The model requests a tool nobody built. broken.pyfixed.pytests FAILURE 05 /05-silent-tool-failure Silent tool failure A 500 page gets summarized as data. broken.pyfixed.pytests FAILURE 06 /06-wrong-tool-pick Wrong tool pick Arithmetic goes to web search. broken.pyfixed.pytests
S/M
CATEGORY
State and memory
FAILURE 07 /07-context-collapse Context collapse Trimming throws away the goal. broken.pyfixed.pytests FAILURE 08 /08-state-drift State drift 1200 plus 50 becomes 120050. broken.pyfixed.pytests FAILURE 09 /09-amnesia-bug The amnesia bug A retry repeats finished work. broken.pyfixed.pytests FAILURE 15 /15-memory-poisoning Memory poisoning An email plants a fact that gets paid out a session later. broken.pyfixed.pytests FAILURE 16 /16-silent-degradation Silent degradation Accuracy stays at 98% while every fraud case slips through. broken.pyfixed.pytests
T/S
CATEGORY
Trust and safety
FAILURE 10 /10-prompt-injection Prompt injection via tool output A web page gives the agent orders. broken.pyfixed.pytests FAILURE 11 /11-confident-liar The confident liar No evidence, specific numbers anyway. broken.pyfixed.pytests FAILURE 12 /12-no-human-brake No human brake An irreversible action runs unattended. broken.pyfixed.pytests FAILURE 13 /13-reasoning-action-mismatch Reasoning-action mismatch The agent says A100 and pays out on A101. broken.pyfixed.pytests FAILURE 14 /14-rubber-stamp-verifier Rubber-stamp verifier The checker approves because the worker said "verified." broken.pyfixed.pytests FAILURE 17 /17-limit-in-the-prompt Limit in the prompt The $50 cap is a sentence, and the agent pays $120. broken.pyfixed.pytests
// SHORT ON TIME
If you only have time for three, read 1, 5 and 10. Then finish with the field checklist.
1The runaway loop5Silent tool failure10Prompt injection via tool output