Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Web agents can succeed in a browser demo and still fail to finish the same kind of work reliably across changing pages and multi-step tasks. That execution gap is visible in benchmark results: WebArena’s 2024 evaluation reported 14.41% end-to-end success for its best GPT-4-based agent, versus 78.24% for human performance on that benchmark. Those figures describe one evaluation, not every live website or today’s models.
The case for infrastructure redesign is an architectural argument, not a proven rule that infrastructure always beats prompt engineering. Benchmarks show the difficulty of dependable end-to-end browser work; Lei Lei’s proposal is to make page state, valid actions, and task context more explicit so agents can execute with less ambiguity.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
What the execution gap means for web agents
The execution gap is the difference between demonstrating a browser workflow under controlled conditions and completing a multi-step task dependably when page layouts, available controls, or task state vary. A model may understand the requested outcome yet still choose an invalid control, miss information in a cluttered page representation, or lose track of what has already happened.
Reliable web automation therefore involves more than generating a plausible next click. It must interpret the current interface, select an action that is valid in that state, observe the result, and carry relevant state forward until the requested outcome is actually reached.
#1 Best Overall
What the benchmark results establish
WebArena: long-horizon tasks across functional websites
WebArena was designed as a realistic, reproducible environment with functional websites in four areas: e-commerce, social forum discussions, collaborative software development, and content management. Its tasks were diverse and long-horizon, making it a test of completing workflows rather than merely identifying a page element.
In the authors’ 2024 evaluation, the best GPT-4-based agent achieved 14.41% end-to-end task success; human performance was 78.24% on the same benchmark. The comparison shows a substantial gap in that environment. It does not predict success rates on every production site, represent all current models, or isolate whether the main cause was the model, browser interface, task design, or execution stack.
WorkArena: enterprise knowledge work
WorkArena evaluates browser agents on knowledge-work tasks based on ServiceNow. The authors describe a 33-task benchmark and report that full task automation remains a considerable gap. Their abstract states: “Our empirical evaluation reveals that while current agents show promise on WorkArena, there remains a considerable gap towards achieving full task automation.”
The WorkArena authors also introduced BrowserGym, an environment for designing and evaluating agents that provides a range of actions and multimodal observations. This matters because browser-agent results depend on how the environment represents a page and what actions it permits, not just on the language model or its prompt.
How to read the numbers
- WebArena’s 14.41% and 78.24% figures are end-to-end results from its 2024 benchmark evaluation, not live-site guarantees.
- WorkArena’s 33 tasks describe its benchmark scope; that count is not a success-rate comparison with WebArena.
- The benchmarks cover different domains and task designs. Their figures should not be used to rank agents against one another unless conditions and task scope are comparable.
Why prompt changes may not close the gap
Prompt engineering can help when a particular failure has a stable, identifiable cause—for example, clarifying how to handle a known form field or specifying a decision rule. But a prompt alone cannot ensure that the page still exposes the same controls, that an action remains valid after a state change, or that the agent retains the correct task context across a long workflow.
Lei Lei’s September 8, 2026 article identifies three interface-level failure modes: noisy page representations, ambiguous action choices, and loss of context across steps. The article recounts a multi-step form test in which a changing production site affected success while the prompt and model remained the same. That account is the author’s experience, not an independently verified controlled comparison.
The distinction is important: prompt improvements may address a specific failure on a stable page, while interface drift and multi-step state management call for support outside the prompt. The benchmark papers establish that end-to-end browser tasks remain challenging; they do not test whether infrastructure redesign as a complete package outperforms prompt engineering.
Recommended Free Tools
What infrastructure redesign would change
Lei Lei’s proposal is to add an intermediate browser layer that makes the agent’s working context more structured and actionable. It has three parts:
Normalize the page representation
Instead of asking the agent to reason over a noisy or inconsistent view of a page, provide a normalized, machine-readable representation of relevant page information. The goal is to make current state easier to interpret, not to assume that every site can be reduced to an identical interface.
Constrain actions to what is valid now
Expose a limited set of actions that are valid for the current page state. This can reduce ambiguity between possible clicks or inputs and make it easier to detect when the interface has changed and the previous action is no longer appropriate.
Preserve task state across steps
Maintain a structured record of the task’s progress and relevant context. A multi-step workflow should not depend on the model reconstructing from scratch what it has already done, what remains, and which results must be checked.
Together, these are design recommendations from Lei Lei’s article, not features whose combined effect has been demonstrated in the cited benchmark studies. They describe a way to target interface interpretation, action selection, and continuity—the points where an otherwise capable agent can lose execution reliability.
Best Value
How to evaluate reliable web automation
A useful evaluation should measure completed outcomes in the environment where the agent is expected to work, not just whether it can select plausible actions in a static example. When reviewing an agent or an automation design, ask:
- What environment and domain? A ServiceNow workflow, an e-commerce task, and a content-management task may expose different controls and failure modes.
- How long is the task? A single action does not test whether state and context survive through a long workflow.
- Is the result end-to-end? Component capability or a successful intermediate action is not the same as finishing and verifying the user’s requested outcome.
- Can it detect changed page state? A reliable system should not blindly repeat an action when the interface no longer matches its assumptions.
- Can it recover? Evaluation should consider whether the agent can re-observe, choose a valid next step, and continue after a page or task state changes.
- Are comparisons like-for-like? Results from different benchmarks, task scopes, and execution conditions do not establish a meaningful ranking on their own.
What the evidence does—and does not—support
The benchmark findings support a measured conclusion: completing complex browser tasks remains difficult in the reported settings. Lei Lei’s article offers a reasoned infrastructure thesis for addressing sources of execution failure that prompts cannot reliably control, especially changing interfaces and context across steps.
They do not establish that prompt engineering is useless, that infrastructure redesign is universally superior, or that one architecture will solve full task automation. The practical implication is narrower: treat the prompt as one part of the system, and evaluate the browser interface layer, action constraints, state handling, and recovery behavior as part of reliable web automation.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesQuick Recap
Sources
- WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?, PMLR, 2024.
- WebArena: A Realistic Web Environment for Building Autonomous Agents, 2024 evaluation paper.
- Lei Lei, “The Execution Gap in Agentic AI,” September 8, 2026.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
