ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-19

Browser Use Explains Browser Agents: Architecture, Selection Boundaries, and Cost Risks

Browser Use published a technical guide explaining that browser agents consist of three layers: model, harness, and browser, operating through an observe-select-execute-check loop. The article compares the use cases for APIs, scripts, and agents, cites internal cloud data on high-frequency tasks like login and extraction, and outlines key limitations including inconsistent results, website blocks, prompt injection, and unpredictable costs.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

Verified 2026-10-19 18:57 GMT+8

AI browser agents are becoming a new tool for automating complex web tasks. Nik Tomazic from the Browser Use team published a detailed technical guide on October 18, 2026, explaining how these agents work internally, when they are appropriate, and what risks they entail. For developers, understanding these details helps decide when to use an agent versus sticking with traditional APIs or scripts.

Architecturally, a typical browser agent consists of three core components: a model (for decision-making), a harness (providing context, tools, and the execution loop), and a browser (maintaining the website session). The model chooses the next action based on the current page state; the harness ensures the 'observe-select-execute-check' loop continues, stopping when time, step, or cost limits are reached; and the browser loads pages and manages tabs/cookies via protocols like Chrome DevTools Protocol (CDP) or WebDriver BiDi. Agents read pages in two main ways: parsing the DOM or accessibility tree for structured text, or using vision models to analyze screenshots for non-standard interface elements.

In practice, developers face a choice between three automation paths: API, script, or agent. Tomazic advises prioritizing an API if the service provides one that supports the task, as it is direct and stable. If no suitable API exists but the workflow is predictable (e.g., downloading the same report every Monday), write a script using Playwright or Selenium, because scripts follow deterministic rules, are low-cost, and easy to debug. Only introduce an agent when the workflow is unpredictable, requires flexible judgment, or exact steps cannot be defined. Agent decisions are stochastic, meaning the same task may take different paths—this is their advantage in adapting to unknown pages but also the root of their inconsistency.

Sign in / Sign up
ReadingResearchRadarInvestment framework

To illustrate real-world usage, Browser Use cited run data from its cloud platform's V2, V3, and V4 agents as of an October 18, 2026 snapshot. The data shows that 'getting past a login' (including one-time codes and MFA) accounted for 18% of runs, followed by 'extracting structured data' (14%), 'researching open questions' (6%), 'applying for jobs' (6%), 'finding filings and reports' (5%), and 'booking and buying' (3%). Notably, these figures describe what users run, not how often tasks succeed.

Despite their power, browser agents have clear limitations. First, websites can actively block agents via bot-management systems or rate limits; a better model won't resolve all access restrictions. Second, expired sessions and multi-factor authentication often require human intervention. Third, agents can produce incorrect or incomplete results, such as selecting the wrong product or missing fields, and retries may lead to duplicate orders. Fourth, prompt injection is a serious security risk where malicious page content might redirect the agent away from the original task. Finally, because they rely on LLMs, running costs are hard to predict, and failed retries further increase expenses.

Addressing these challenges, Tomazic proposed a cost accounting method: when comparing solutions, divide the total spend across all attempts (including failures and retries) by the number of verified successful tasks, tracking human review time separately. He suggested starting with read-only tasks (like collecting product prices), clearly defining allowed pages, required fields, and success criteria, and setting strict step or cost limits. Only expand to more complex sites or tasks after understanding where the agent needs help.

Sources

  • 原始出处browser-use.com↗
Continue researching →