AI ranking

Browser agents

Agents that drive a browser (OpenClaw browser, Manus browser, autonomous web agents). Assessed on navigation, screen observation and verification of the result.

Ranking updated on 02 November 2026 at 11:47

Data temporarily unavailable. The ranking is still loading from the live data. If the problem persists, check your connection or go to the main ranking page.

Updated automatically about every 25 hours (BENCH GitHub Actions → FTP /data/bench/).

Le Recul comparator

Who wins on what

Tick at least two models in the ranking table, then come back here.

Indicative score, weighted from public sources. Transparent methodology. See the AI ranking page for the full methodology and the other categories.

AI browser agent ranking

Le Recul's Browser agents ranking covers agents that drive a web browser: they read the page, click, fill in forms and chain steps to complete an online task on your behalf.

This is the most exposed family of agents. A browser agent acts on real sites, often connected to your accounts, and the web is not a cooperative environment: pages change, content tries to divert the agent from its instruction, and some actions are hard to reverse. The assessment therefore bears as much on caution as on capability.

Public coverage of this category remains limited and the entries are few. Read the ranking as a cautious reference point on a field that is still young, not as an established hierarchy.

What this ranking is for

This ranking helps you understand what a browser agent can really do, beyond demonstrations where the target site behaves perfectly.

Above all it is a reminder of the precautions to take: limit the accounts it can reach, avoid irreversible actions, and never hand credential entry or payment to an agent.

How to read the ranking

  • Web tasks actually carried through, not merely started.
  • Reading the screen and understanding the state of the page.
  • Checking the result obtained before carrying on.
  • Control over permissions and sensitive actions.
  • Resistance to page content trying to divert the instruction.

FAQ — Browser agents

Can a browser agent make purchases for me?

Technically some offer it, but that is the kind of action to avoid delegating: it commits money and is hard to undo. Payment and credential entry should stay manual.

Why are there so few entries in this ranking?

Because the category is young and comparable public evaluations are rare. A few documented entries are better than a padded list with no measurement behind it.

What is a prompt injection from a page?

Text placed in a web page that addresses the agent to divert it from its task. Resistance to that kind of content is a central safety criterion for this category.

Should you watch the agent while it works?

Yes. On committing actions, supervision remains necessary: it is the only way to interrupt a sequence going wrong.

How often is the ranking updated?

About every 25 hours, automatically. The update date is shown at the top of the page.

See the full FAQ →

Related rankings