Diagnosing with Insights: Structured Analysis of Agent Failures via Behavioral Abstractions

2609.02371 ↗ · 2026-09-02 · cs.AI

Jiayi Bi, Yanjie Gao, Yuanmin Xie, Liqun Li, Tianyin Xu, Fan Yang, Mao Yang

The finding

Headline claim

A neuro-symbolic framework (AgentScope) that abstracts agent trajectories into structured graphs and checks LLM-judged 'neural invariants' localizes and classifies agent failure causes substantially more accurately than raw LLM-as-a-judge diagnosis.

The test

Failure localization and attribution on Who&When (126 algorithm-generated plus 58 hand-crafted annotated failure trajectories) and on AgentErrata, the authors' new fault-injection benchmark of 11 failure-mode types, using GPT-4o, GPT-5.1, and DeepSeek-V3.2 as underlying judges.

Effect size & scope

Best vanilla LLM judge (GPT-5.1) reaches only 18.15% attribution accuracy; AgentScope spans 25.40-77.78% localization on Who&When Algo-Generated, 22.41-34.48% on Hand-Crafted, and 28.38-54.13% on AgentErrata. Absolute accuracies remain low; the wide ranges suggest high variance across conditions, and the headline gain partly rests on the authors' own fault-injected benchmark.

Limitations

Authors concede the Hand-Crafted subset is small (58 traces) and that Who&When lacks failure-type labels, motivating their self-built dataset whose injected failures are designed to match their own taxonomy, a built-in advantage. Unstated: neural invariants are themselves LLM judges with structured prompts, so the framework may inherit prompt sensitivity; no runtime overhead numbers appear in the provided text; baselines are all pure-LLM judges, not other structured diagnosis methods.

Counterpoint

The comparison may be rigged in the framework's favor: AgentErrata failures were injected according to AgentScope's own 11-mode taxonomy, so the classifier is being graded on its own answer key, and the 54.13% figure may not transfer to organic failures. Also, the decisive-judgment stage is still an LLM call, so 'deterministic' framing overstates how symbolic the system really is.

Economy implication

Little or nothing yet for the AI economy in the next 12 to 24 months: this is debugging tooling research, though if agent failure diagnosis matures it is a precondition for the unattended reliability that enterprise ROI claims depend on.

Who this matters for

Technical Capability: Provides a concrete measure of how bad current LLM-judge failure attribution actually is (18.15% for GPT-5.1), a useful calibration for anyone trusting agentic self-evaluation.

Market & Valuation: Weakly relevant: reliable agent observability and diagnosis tooling is an emerging enterprise spend category adjacent to agent deployment.

Bears on (advisory · papers never write evidence)

[1.3] Agentic reliability (multi-step completion without human correction) is improving materially, not just peak single-shot.

Abstract

With the proliferation of LLM agents, the ability to understand and diagnose failures in agents is essential to achieving superior effectiveness and trustworthiness. As agent failures often manifest via long and complex trajectories, manually finding the needles in the haystack is untenable. However, traditional diagnosis techniques for software bugs can hardly address LLM agent failures, while completely relying on LLMs as the judge yields unreliable diagnosis results. To overcome these challenges, this paper presents AGENTSCOPE, a new neuro-symbolic approach for agent failure mode diagnosis. The key principle of AGENTSCOPE is to abstract agent behavior, based on its trajectories, into structured representations. Furthermore, AGENTSCOPE introduces the concept of neural invariants to specify agent behavior properties. AGENTSCOPE leverages LLM-guided reasoning atop the structured representation against neural invariants to pinpoint both the failure step and its type in the trajectory. We show the effectiveness of AGENTSCOPE on publicly available agent failure datasets (Who&When) and a more comprehensive dataset created by us (AgentErrata), where AGENTSCOPE significantly outperforms the current state of the art in fault localization and attribution accuracy. Our work shows that integrating structured abstractions with LLM-guided reasoning enables effective, reliable, and interpretable diagnosis for agent failures.