A taxonomy of autonomous agent failure modes

Thesis

A useful failure taxonomy connects observed agent behavior failures to the specific instrumentation that would have caught them.

Outline

1. 01
   Planning failures

   Goal misinterpretation, decomposition errors, and ignored alternatives.

2. 02
   Tool-use failures

   Wrong tool, wrong inputs, unbounded permissions, partial execution.

3. 03
   Memory failures

   Stale, irrelevant, or over-weighted retrieved context.

4. 04
   Coordination failures

   Multi-agent and human-in-the-loop handoff breakdowns.

5. 05
   Observability mapping

   Which trace surface catches each class of failure.

Status