# Laminar — Taxonomy of agent failures: how AI agents fail in production

- Company: Laminar (laminar.sh)
- Announced: 2026-11-10
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://laminar.sh/blog/taxonomy-of-agent-failures
- Record: https://forck.live/items/34197-taxonomy-of-agent-failures-how-ai-agents-fail-in-production
- Subject: flow / agent observability
- Models affected: claude-sonnet-5, gpt-5.6-sol, gpt-5.4-mini, claude-opus-5-5, gpt-6.1-sol, gpt-5.6-luna, claude-haiku-4-5, gpt-4.1, claude-sonnet-5-5, claude-opus-4-6

Laminar sorted thousands of failures from production agents into 7 categories plus an "Other" bucket and published the resulting taxonomy, after running its Signals trace intelligence on agent traces from coding, computer use, legal and finance. Hallucinations account for 26% of failures, with plain fabrication making up about half of that category and nearly 2 in 3 agents fabricating at least once. Unfinished tasks are 21% of failures and appear in 9 in 10 agents, often as traces that simply stop or hand the task back to a human without exhausting the agent's own tools. Breaking an explicit instruction is 20% of failures, most often a skipped mandatory step, a content or style violation, or malformed output such as JSON wrapped in Markdown fences. Tool errors, 12% of failures, are the kind agents recover from most: more than a third are fixed because the errors come back as messages. Wasted work, 9% of failures, is mostly loops and retry storms, and Laminar says open-weight models loop 1.7x as often as other agents of the same type, at 1.7x the tokens of a typical flagged run. The "Other" bucket covers infrastructure failures, defects in the agent's own prompt or input data, and multi-agent hand-off issues; only 13% of failures are recovered even partially, and agent type predicts the failure mix about 1.5x as much as the model vendor.

## Evidence

Verbatim from https://laminar.sh/blog/taxonomy-of-agent-failures:

> Laminar runs Signals, our trace intelligence, on agent traces from coding, computer use, legal, finance and other domains. Each trace is analysed in isolation, and the findings are clustered into recurring patterns. Learn more in the Signals docs .
> 
> How do agents actually fail? Do the patterns repeat across industries? Do failures depend on the harness, the model provider, or the model size?
> 
> We sorted thousands of failures from production agents into 7 categories, plus "Other".
> 
> TL;DR #
> 
> "Other" covers infrastructure failures (failed LLM calls, context-window overflows, blocked web pages), defects in the agent's own prompt or input data, and multi-agent hand-off issues.
> 
> Almost every agent fails in more than one way. 9 in 10 agents left at least one task unfinished, and more than 4 in 5 hallucinated at least once.
> 
> Fabrication (inventing facts, numbers or citations) is the most common single failure type. It makes up about half of the hallucination category. Nearly 2 in 3 agents do it at least once.
> 
> Agents rarely recover on their own. Only 13% of failures are recovered, even partially.
> 
> Agent type predicts failures better than the model does.

---

Record: https://forck.live/items/34197-taxonomy-of-agent-failures-how-ai-agents-fail-in-production
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md
