---
title: "Every failure, automatically classified — Retrace"
description: "Auto-classify failed AI agent runs into 14 standard MAST failure modes, surfaced as detections and an Insights breakdown of why agents fail."
canonical: "https://retraceai.tech/features/failure-taxonomy"
last-updated: 2026-08-26
format: text/markdown
---

# Every failure, automatically classified

When a run fails, Retrace automatically sorts it into one of 14 standard agent failure modes, so you learn why your agents fail across the whole project, not just that they did.

## What it does

- **14 failure modes** — Failed runs are labeled with a mode from a standard multi-agent failure taxonomy (MAST), so failures are comparable across runs and projects.
- **Two-stage classification** — A fast semantic pass narrows the likely modes and a judge confirms, accurate without paying to deeply judge every run.
- **Failures only** — Classification runs only on runs that actually failed, so cost is bounded to the population that matters.
- **Metered separately** — Classification uses a dedicated allowance kept apart from your interactive AI usage, and runs once per failed run, it never double-charges or starves your quota.
- **Surfaced everywhere** — Each classification appears as a detection on the run and rolls up into an Insights breakdown and a trend over time.
- **Safe by default** — If a confident classification can't be produced, the run is left unlabeled, never an error, and never a wrong label.

## REST API

- `GET /api/v1/insights/failures: failure breakdown`
- `GET /api/v1/insights/drift: drift signals`
- `GET /api/v1/insights/anomalies: anomaly signals`
- `GET /api/v1/detections: detections on your runs`

## Related

- HTML page: https://retraceai.tech/features/failure-taxonomy
- Documentation: https://docs.retraceai.tech
- OpenAPI specification: https://retraceai.tech/openapi.json
- Authentication for agents: https://retraceai.tech/auth.md