Six failure modes in tool-using agents, and the patterns that fix them
An empirical taxonomy of agent tool-use failures across 4,000 traces from production deployments. Schema drift and silent partial-failure dominate.
A taxonomy of agent tool-use failures derived from 4,000 anonymized production traces. Two modes account for 63% of incidents: schema drift (tool definitions silently change between deploys) and silent partial-failure (tool returns success with degraded data).
What changed. A clean failure taxonomy with empirical frequencies, instead of anecdotes.
Why it matters. Most agent post-mortems blame the model. The data says most agent incidents are caused by tools, not the planner.
Builder takeaway. Wrap every external tool with a contract test that runs in CI. Add a result validator that asserts shape and freshness, not just status code.