The 14 ways AI agents fail in 1,642 traces and the security failure OWASP names

A study of 1,642 runs of multi-agent systems found 14 ways they fail, in three groups: design, agents not coordinating, and weak checking of the result. OWASP names a separate failure: an agent allowed to do too much. Read on 2026-11-06.

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

Contents

Two bodies of work give a list of how AI agents fail. A research group at UC Berkeley and collaborators built a catalogue of 14 failure modes from 1,642 recorded runs of multi-agent systems built on seven frameworks. A security community, OWASP, describes a different failure: an agent that is given more power than its task needs. They cover different ground: the first is about why systems do not finish or do the wrong task, the second is about what an agent can damage when something goes wrong.

The study found 14 failure modes in three groups

The paper "Why Do Multi-Agent LLM Systems Fail?" (submitted March 2025, revised October 2025) says that "their performance gains on popular benchmarks are often minimal."[1] The authors assembled a dataset of "1600+ annotated traces collected across 7 popular MAS frameworks", built a taxonomy from 150 of them with expert human annotators ("kappa = 0.88"), and the process "identifies 14 unique modes, clustered into 3 categories: (i) system design issues, (ii) inter-agent misalignment, and (iii) task verification."[1] The full paper reports the prevalence of each mode as observed across 1,642 traces.[2]

Two findings matter for a buyer. First, systems with a verification step failed less: "Systems with explicit verifiers like MetaGPT and ChatDev generally show fewer total failures", though "the presence of a verifier is not a silver bullet, as overall MAS success rates can still be low."[2] Second, changing the instructions given to one agent helped: "improving agent role specifications alone yields a +9.4% success rate increase for ChatDev with the same user prompt and LLM (GPT-4o)."[2]

The scope is narrow. The systems studied are frameworks in which several language-model agents talk to each other, mostly on coding, maths and general tasks, and the models tested were GPT-4, Claude 3, Qwen2.5 and CodeLlama.[1] The percentages are prevalence figures from the paper's own annotation of its own collected runs.[2]

The security failure is excessive agency

OWASP, an open security community, lists "Excessive Agency" among the risks for applications built on language models: "the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM."[3] It names three root causes, excessive functionality, excessive permissions and excessive autonomy, and recommends that you "require a human to approve high-impact actions before they are taken."[3] This is a design standard, not a count of how often it happens: the page measures no rate.

Where the other tests show the same failures

Synthesis, this page's own mapping and not any source's claim. Penrose's AccountingBench reports that models "accumulate material errors over time" while closing books month after month, which looks like the study's loss-of-context and weak-verification modes.[4] The accounting benchmarks, and what job projections say about bookkeepers, are on Will AI replace bookkeepers and accountants?. IBM warns that an agent with trouble planning or tracking progress "may get stuck repeating the same actions or fail to complete its assigned task", which is the study's step-repetition mode.[5] Whether the same causes are behind them is not shown by any of these sources. How Anthropic separates what an agent says from the state it leaves behind, and how often retrieval-backed tools still hallucinate, is on Testing AI agents and hallucination rates.

Not yet proven

If you run an agent that has failed in a way not listed here, I would like to compare notes: mackenzie@mackenziebowes.com.

Sources

  1. Cemri et al. (UC Berkeley and collaborators), Why Do Multi-Agent LLM Systems Fail?, arXiv 2503.13657, submitted 2025-03-17, revised 2025-10-26 (abstract page). arxiv.org/abs/2503.13657
  2. Cemri et al., Why Do Multi-Agent LLM Systems Fail?, arXiv 2503.13657v3 (HTML version of the full paper, sections 4 and 5 and appendix). arxiv.org/html/2503.13657v3
  3. OWASP Gen AI Security Project, LLM06:2025 Excessive Agency (page dated 2025-05-05 in the search result). genai.owasp.org/llmrisk/llm062025-excessive-agency/
  4. Penrose, AccountingBench: Evaluating LLMs on Real Long-Horizon Business Tasks (2025; the page carries no date). accounting.penrose.com/
  5. IBM, AI Agents vs. AI Assistants, IBM Think, read 2026-11-06, undated. www.ibm.com/think/topics/ai-agents-vs-ai-assistants

Terms on this page

  • AI agent: A system where a language model decides for itself how to do a task and which tools to use.
  • Multi-agent system: Several AI agents working together, each handling part of a problem.