AI agents

An agent arrives looking like a browser, because it usually is one. Computer use works from screenshots and pixel coordinates, moving a cursor and typing the way a person does^[2], and Atlas gives its agent mode a visible cursor inside the ordinary browser window^[4]. Claude for Chrome runs as an extension taking actions in the user's own browser^[5]. That leaves a site with no reliable string to match on, which is why the identification work moved to signatures rather than user agents. See AI crawlers.

Web Bot Auth proposes that bots sign requests cryptographically, using RFC 9421 HTTP Message Signatures, so a site owner can verify who is calling rather than trust a header^[6]. Cloudflare put it into production on 28 August 2025 with ChatGPT agent, Goose from Block, Browserbase and Anchor Browser as the first signed cohort^[7], and now verifies a bot only where it declares itself through a signature, a published IP list with a stable user agent or reverse DNS, and obeys robots.txt at reasonable rates^[8]. The standards work is unfinished: the original architecture draft expired in March 2026 and was replaced by a protocol draft^[9].

Public panels do not yet break agents out. A vendor crawler panel splitting AI bot requests by declared purpose for January to July 2026 has rows for training at 46.5%, mixed purpose at 39.3%, search or answer retrieval at 11.1% and user-triggered action at 2.5%, and none for agent activity^[10], so agent traffic sits inside the user-triggered slice or inside undeclared browser traffic. The published measurements so far are safety rather than volume: Anthropic reported prompt-injection attacks succeeding 23.6% of the time against browser use without mitigations and 11.2% with them^[11], and browser-specific attacks falling from 35.7% to 0%^[12], on a pilot opened to 1,000 users^[5].

Benchmarks put agent reliability far below a person's. WebArena, a reproducible environment built from fully functional websites in four common domains, reports its best GPT-4-based agent completing 14.41% of tasks end to end^[13], against 78.24% for human testers on the same tasks^[14]. The gap has narrowed rather than closed: Branch-and-Browse, a tree-structured browsing agent published in October 2025, reaches 35.8% on WebArena^[15], under half the human rate. The task sets are drawn from real sites rather than mock-ups, which is what makes the numbers transferable: Mind2Web collects over 2,000 open-ended tasks from 137 websites spanning 31 domains^[16].