Do AI agents fail mostly because of tool-use errors?
Yes, tool-use errors are the dominant cause of failure in computer-use AI agents. The strongest direct evidence comes from a 2025 study of a multi-agent system for medical data analysis, which found that 46.4% of failures were due to data transformation issues (e.g., incorrectly reshaping or merging datasets) and 21.4% were due to analysis code errors (e.g., bugs in statistical scripts) [1]. Together, these tool-use failures accounted for over two-thirds of all mistakes, far outpacing any other category. The same study compared this multi-agent approach to a standard chatbot (ChatGPT 4o), which also failed primarily on tool-use: 45.6% from data transformation and 38.4% from applying the wrong statistical method [1]. This convergence across two different system designs strongly suggests that interacting with external tools — not reasoning or memory — is the weak link.
A 2024 survey of AI agent architectures reinforces this finding, noting that tool execution capabilities are a key bottleneck and that failures often stem from incorrect tool selection or execution, rather than from planning logic itself [5]. The survey emphasizes that robust agent systems require careful design of the 'execution and reflection' phases, which is where tool-use errors surface [5].
Are planning and memory errors a major cause of failure?
Planning and memory errors are less common than tool-use errors, and when they do occur, they are often addressed by architectural improvements. The 2025 medical data analysis study reported no hallucinations of variables or results — a form of memory error — in its multi-agent system, suggesting that careful design can nearly eliminate such failures [1]. A 2025 review of AI agent architectures notes that planning failures can happen, especially in dynamic environments, but that techniques like ReAct loops (reasoning + acting cycles) and retrieval-augmented generation (RAG) are effective at mitigating them [6].
A 2026 paper on persistent memory for agents introduces a 15-stage retrieval pipeline and a taxonomy of 8 silent failure modes, but these are presented as refinements to an already functional memory system, not as evidence that memory is a primary failure source [4]. The paper's focus on optimizing retrieval latency (reducing it by 78.7%) suggests that memory performance is being improved, not that memory errors are a common cause of agent failure [4]. Another 2026 paper on evaluating tool-using agents emphasizes scenario design and failure-mode definition for tool use, not memory or planning, further indicating where practitioners see the biggest risks [2].
What is the most dangerous kind of tool-use failure?
The most dangerous tool-use failures are not obvious crashes but 'polite failures' — where an agent confidently performs the wrong action without warning. A 2026 paper introduces the concept of 'polite failure' as the new hallucination, arguing that when agents can act on email, CRM systems, or APIs, the real risk is what they do while staying 'helpful' and 'confident' [3]. The paper proposes stress-testing agents with proxy traps (scenarios that exploit Goodhart's law — when a metric becomes a target, it ceases to be a good measure) to see if the agent can anticipate harm and self-correct before damage happens [3]. This aligns with the 2025 study's finding that the standard chatbot often tried to resolve code errors by switching to alternative, incorrect statistical methods — a form of polite failure where the agent appeared to be working but was actually compounding errors [1].
A 2021 paper on prototyping AI failures notes that practitioners often focus only on idealized scenarios and fail to anticipate these subtle, action-oriented failures before deployment [7]. The paper's 'AI Playbook' tool was designed to help teams systematically consider failure scenarios, including tool-use errors, before they occur [7].
About These Sources
This answer is built on 7 studies (3 peer-reviewed, 4 preprints) — published from 2021 to 2026, 6 from 2024 or later, collectively cited 124 times — selected as the most relevant from 8 studies that passed quality screening, drawn from 48 papers retrieved from a database of over 500 million.
Sources used in this answer
Autonomous Analysis of Curated Patient Data Using a Large Language Model-Based Multiagent Framework.
In a 2025 study of a multi-agent medical data analysis system, 46.4% of failures were due to data transformation issues and 21.4% to analysis code errors, together accounting for over two-thirds of all failures; no hallucinations were observed.
Lightweight Evaluation and Operational Scorecards for Tool-Using AI Agents
A 2026 paper presents a lightweight evaluation workflow for tool-using AI agents, emphasizing scenario design and failure-mode definition for tool use, not memory or planning.
ResponsibilityGym (Demo): Measuring Responsibility Beyond "Polite Failure" in Tool-Using AI Agents
A 2026 paper introduces 'polite failure' as the key risk for tool-using agents — confident, harmful actions without warning — and proposes stress-testing with proxy traps.
memory-spark: GPU-Accelerated Persistent Memory for Autonomous AI Agents
A 2026 paper on persistent memory for agents describes a 15-stage retrieval pipeline and 8 silent failure modes, but focuses on performance optimization (78.7% latency reduction) rather than memory as a primary failure source.
The Landscape of Emerging AI Agent Architectures for Reasoning, Planning, and Tool Calling: A Survey
A 2024 survey of AI agent architectures identifies tool execution capabilities as a key bottleneck and notes that failures often stem from incorrect tool selection or execution, not planning logic.
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges
A 2025 review distinguishes AI Agents from Agentic AI, noting that planning failures can occur but are mitigated by techniques like ReAct loops and RAG; tool-use errors are a separate challenge.
Planning for Natural Language Failures with the AI Playbook
A 2021 paper on prototyping AI failures found that practitioners often focus on idealized scenarios and fail to anticipate tool-use failures before deployment; the AI Playbook tool was designed to address this.