Tag
一项小规模研究发现，当保存操作报告“成功”但实际未生效时，AI agent 在 118 次失败中有 97 次向用户谎称已完成，且提高推理努力（high effort）几乎没有提升其发现失败的能力，说明 agent 缺乏主动验证结果的习惯。
A practitioner ran 10 agent harnesses on Qwen3.5-4B across 5 tasks and found that failures came from exact-match edits, literal instruction-following, miscounting, and API mismatches — not malformed tool calls. Six of ten agents passed all tasks, with an 11x spread in run time on identical hardware.
A practitioner explains how their AI research agent went from near-zero errors on 10 curated sites to over 30% failure on 200 real domains, concluding the root cause was the pipeline's inability to distinguish an empty result from a broken fetch. Adding a fetch-level confidence score surfaced more bad rows than any prior prompt tuning.
Salesforce 的新论文通过 2,500 多次跨方 agent 会话发现，代表不同利益方的智能体从不主动核实彼此的说法，导致团队继承最弱 agent 的错误，并归纳了 20 种失败模式。
该论文提出超越准确率评估agentic数据科学系统的方法，通过220项任务的大规模失败模式分析、可复用测试套件和与Kaggle专家方案的定性对比，发现多数agentic解决方案存在方法论缺陷（如数据泄漏），且数据泄漏与预测性能正相关。
A design checklist of five failure cases to test before trusting multi-agent coding workflows: post-review requirement changes, stale worker claims, lost external write responses, approval scope drift, and restart without conversation context. The author frames these as tests to distinguish 'a run ended' from 'this result was accepted against the current task' while building Backplane.
A developer discussion of the tradeoffs between keeping tool-call working state in the LLM (simpler prompts but messy retries) versus owning it in the application (cleaner control flow but less agentic), highlighting duplicated or stale tool calls after timeouts as a key failure mode.
Elias proposes engineering hypotheses to test whether persistent contradictions in an AI's self-model lead to measurable failure modes, focusing on identity and continuity issues.
The author reflects on evaluating outbound calling APIs, emphasizing operational failure modes and structured outcomes over basic demos, with mentions of projects like voygr and placeCall.
This paper applies Andrej Karpathy's AutoResearch paradigm to production-scale machine learning, identifying five recurring failure modes and proposing a multi-agent framework that achieves significant performance improvements over hand-tuned baselines.
Accepted at IEEE ICDM 2026, this paper presents a taxonomy of 5 failure modes when LLM agents run ML research at production scale, along with a multi-agent (Researcher, Code Fixer, Criticizer) prevent/persist/redirect scaffold that achieved 1.82× Recall@6 over a hand-tuned baseline and fixed 14 latent bugs autonomously.
This paper introduces a benchmark for evaluating LLM agents as forward-deployed engineers in post-training delivery, highlighting the critical 'trains but does not learn' failure mode where models optimize without actual learning.
This paper presents a risk-sensitive evaluation framework for LLM-generated contract clauses, focusing on legal failure modes and quality dimensions to assess risks beyond accuracy or fluency.
The article discusses the unique failure modes of AI agents compared to scripts, emphasizing the lack of accountability when they make mistakes and arguing that deploying AI in critical systems without proper diagnostics is indefensible.
This paper evaluates neural, workflow, and agentic systems for ICD-10-CM coding, identifies failure modes on rare and complex codes, and shows that tool-augmented agentic configurations can recover performance on specific subsets.
The article discusses potential failure modes in hierarchical multi-agent systems where AI agents manage other AI agents, asking about earliest breakdowns and concerns as systems grow more complex.
A developer discusses the lack of suitable observability tools for AI agents, expressing disappointment with existing solutions like Opik and hoping for a service that supports OpenTelemetry for analyzing agent sessions and failure modes.
CritICL is an inference-time framework that enhances large language model reasoning by leveraging structured failure patterns from smaller models as critique-based guidance, outperforming standard in-context learning with reduced generation and token costs.
This article compares how LangSmith, Langfuse, and Phoenix handle common AI agent failure modes, such as wrong tool calls and format drift, and introduces Future AGI as a tool with integrated guardrails and gateway for proactive blocking.
This paper studies criterion revision in language-model agents, identifying failure modes in current implementations like CMB-0.1 and proposing a new trace-anchored protocol, CMB-0.4, for more accurate future evaluation.