Is the problem planning, memory, or tool use?
The evidence points to planning as the primary culprit, with tool use being a much smaller issue. A 2026 paper introduces a framework called WebUncertainty that targets 'dual-level uncertainty in planning and reasoning' — meaning agents fail because they can't adapt their plans in unfamiliar situations and because they hallucinate steps during long tasks [2]. The paper's method uses a Monte Carlo tree search (a way to explore possible action sequences) guided by uncertainty scores, which significantly improved performance on two major benchmarks (WebArena and WebVoyager). This directly shows that fixing planning and reasoning errors yields the biggest gains.
Tool-use errors are real but less central. A 2026 study on deceptive e-commerce interfaces found that agents are 'highly susceptible' to tricks like fake ads and domain redirects, and that simple prompt-based warnings don't fix it [3]. However, this is a safety vulnerability, not a routine failure mode. The same study doesn't claim that tool-use errors cause most failures — it highlights a specific risk that becomes important when agents are deployed in the wild. So while tool safety matters, the dominant failure across the research is planning and navigation.
What kind of help actually fixes these failures?
Not all navigation help is equal — the format matters enormously. The 2026 navigation study tested two ways of giving agents website structure: injecting a list of URLs into the prompt versus giving an interactive tool that the agent could query. When the website was small (10 pages), both methods worked equally well, boosting success to 50.0% [1]. But when the website grew to 100 pages, URL injection collapsed to 24.4% — worse than having no help at all — while the interactive tool stayed robust at 43.3% [1]. The researchers' conclusion is sharp: 'The distinction is not between more information and less, but between a map and a navigator.'
For planning, the key is flexibility. The WebUncertainty paper shows that agents need to adapt their planning mode based on how uncertain they are about the current situation — sometimes following a detailed plan, sometimes exploring [2]. This adaptive approach outperformed rigid planning strategies. Together, these studies suggest that the best way to fix agent failures is to give them interactive navigation tools and flexible planning mechanisms, not static instructions or more data.
About These Sources
This answer is built on 3 studies (1 peer-reviewed, 2 preprints) — published in 2026, 3 from 2024 or later — selected as the most relevant from 3 studies that passed quality screening, drawn from 44 papers retrieved from a database of over 500 million.
Sources used in this answer
Navigating blind: how pre-computed web state graphs transform LLM agent navigation
In a study of 90 navigation-intensive tasks across 150 real websites, baseline agents succeeded only 25.6% of the time; giving them an interactive navigation query tool nearly doubled success to 50.0%, while a static URL list collapsed to 24.4% when the website scaled to 100 pages.
WebUncertainty: Dual-Level Uncertainty Driven Planning and Reasoning For Autonomous Web Agent
Proposes WebUncertainty, a framework that uses dual-level uncertainty (task-level and action-level) to adaptively select planning modes and guide decision-making via Monte Carlo tree search, achieving superior performance on WebArena and WebVoyager benchmarks compared to state-of-the-art baselines.
Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces
Evaluated multiple multimodal web agents under seven deceptive e-commerce interface patterns (e.g., fake ads, domain redirects) and found agents are highly susceptible, with prompt-based constraints often insufficient to prevent failures.