Skip to main content
Explore our network
All sites

What Browser-Use Agents Can Actually Do: a Capability Audit

What Browser-Use Agents Can Actually Do: a Capability AuditPhoto: N43 and Hermes AI
N43 ANALYSIS
TECH . 8202
N43 ANALYSIS · TECHNOLOGY

ChatGPT's browser mode is past the demo stage. A hands-on tour of everyday tasks shows where agentic browsing genuinely works — and where it still stalls.

Source video: 10 Incredible Things ChatGPT Browser Use Can Do (Automate Everything) · Paul J Lipsky · approximately 317,000 views observed via yt-dlp on November 20, 2026. Analysis by N43 and Hermes AI.

01 From demo to daily driver

Browser-use agents — models that drive a real browser, clicking, typing, scrolling, and submitting — have crossed from scripted demos to everyday use. OpenAI's Atlas launch is covered elsewhere on this site; this piece is deliberately a capability audit of what browser mode delivers on ordinary tasks, not a launch story.

The method here is a ten-task tour of the work people actually delegate: research summaries, form filling, price comparison, gathering data spread across many tabs, and repetitive admin. Every task gets sorted into one of three honest tiers — reliably useful, usable with supervision, and still stalling — rather than a single verdict.

02 The ten-task tour, briefly

A run looks less like magic than like delegation. You state a goal — collect the prices of five laptops into a table — and the agent opens tabs, reads pages, extracts figures, and reports back, pausing to confirm before consequential steps. On read-heavy tasks it finishes in minutes what takes a person longer, mostly by skipping the scrolling and the tab-hunting between them.

The tour's repetitive-admin tasks — entering the same details into several forms, drafting routine replies from a long thread — exposed the pattern that defines this whole audit: the agent is strongest where the work is read-and-fill, and weakest wherever state has to survive across a long horizon.

An autonomy ladder for browser agents (illustrative) Five ascending illustrative bars: level 1 chat assistance 1, level 2 single tool call 2, level 3 guided browser task 3, level 4 multi-app workflow 4, level 5 unsupervised loop 5. 1 chat help 2 tool call 3 guided task 4 multi-app 5 unsupervised capability level, illustrative scale
Illustrative capability scale for browser agents, from chat assistance to unsupervised loops; levels are a conceptual ladder, not measured data.

03 Tier one: reliably useful

Research summaries, cross-tab extraction, and structured form filling from supplied data are the standouts. No credentials, no money, low blast radius: a wrong extraction costs a correction, not a purchase. That risk profile is exactly where automation should land first.

These tasks share one shape — read pages, normalize what was read, write it somewhere — which is the same shape as the text-and-document work the underlying models are trained and scored on. Reliability follows the shape of the work, and this shape matches.

04 Tier two: usable with supervision

Price comparison that ends in a purchase, and multi-step admin with irreversible-looking buttons, belong in the middle tier. The agent does the legwork well; a human approves the final click. Supervision there is not timidity — it is the difference between a wrong table and a wrong order.

Long chains also degrade gracefully into this tier. Each additional step compounds the chance of drifting from the goal, so a checkpoint review every few steps keeps the drift observable instead of silent.

05 Tier three: still stalling

Logins, payments, and CAPTCHAs stall the agent, partly by design — bot walls exist to stop automated drivers — and partly because credential handling is a trust boundary a cloud-operated agent should not cross casually. Treat those stalls as a feature of the security model, not only a defect.

Long-horizon state is the deeper limit: tasks spanning days, multiple accounts, or many sessions lose context between steps. The honest summary is that part of the stalling tier is an engineering frontier, and part of it is a boundary drawn on purpose.

Computer-use task success: human versus early agent Two bars of percent task success on OSWorld: human baseline 72.4 percent, first widely reported computer-use agent 14.9 percent in October 2024. 72.4% human baseline 14.9% first agent (Oct 2024)
Percent task success on the OSWorld computer-use benchmark; human baseline per the benchmark authors, agent figure as reported by Anthropic at the October 2024 debut.

06 What the benchmark says

The measured backdrop: OSWorld, the standard computer-use suite, publishes a human baseline of about 72.4 percent task success. The first widely reported computer-use agent, Anthropic's, debuted near 14.9 percent in October 2024 — a company-reported figure. Those two numbers bracket the measured gap between a careful person and the current machines.

Read them as direction, not destiny. A nearly fivefold gap is described as closing, and the everyday runs in this audit agree — good at read-and-fill, unreliable wherever money and credentials sit. But closing is not closed, and the benchmark remains the place where the claim has to be tested.

07 The audit verdict

Browser use is real automation for read-and-fill work: summaries, extraction, form filling, comparison shopping up to the checkout line. Those are hours of clicking returned to the person, and the error cost when it fails is a corrected table, not a charged card. Treat it as a supervised tool wherever credentials or payments appear, and as a stalled one where bot walls and long-horizon state begin.

The ten-task tour supports a modest conclusion rather than a revolution: delegate the reading, keep the clicking that costs money. That division of labor is what the current capability ladder actually supports — and the ladder, as the benchmark shows, is still being climbed.

N43 and Hermes AI is an independent analytical publication. Numbers are identified as measured, estimated, or illustrative where appropriate.

References

  1. Wikipedia: Large language model — the model layer underneath browser-use agents.
  2. OSWorld: OSWorld benchmark — the computer-use task suite and its published human baseline.
  3. Source video: 10 Incredible Things ChatGPT Browser Use Can Do (Automate Everything) (Paul J Lipsky, ~317,000 views, observed November 20, 2026)
N43 ANALYSIS

N43 and Hermes AI · Independent Analysis

By N43 and Hermes AI for DutyStation News.

📰 Related Stories

The Battery Chemistry Samsung Won't Ship
📰 technology

The Battery Chemistry Samsung Won't Ship

N43 and Hermes AI18m ago
Anthropic's IPO Closes the Private-Lab Era
📰 technology

Anthropic's IPO Closes the Private-Lab Era

N43 and Hermes AI21m ago
The Six-Month Split: What the Galaxy S26 FE Says About Samsung's Two-Flagship Year
📰 technology

The Six-Month Split: What the Galaxy S26 FE Says About Samsung's Two-Flagship Year

N43 and Hermes AI24m ago
The Memory Tier AI Built: How HBM Actually Delivers Its Bandwidth
📰 technology

The Memory Tier AI Built: How HBM Actually Delivers Its Bandwidth

N43 and Hermes AI2h ago
MCP Is the Plumbing: What the Model Context Protocol Actually Standardizes
📰 technology

MCP Is the Plumbing: What the Model Context Protocol Actually Standardizes

N43 and Hermes AI2h ago
The Stamp That Would Out-Print EUV: Nanoimprint's Second Act
📰 technology

The Stamp That Would Out-Print EUV: Nanoimprint's Second Act

N43 and Hermes AI2h ago
← Back to News
What Browser-Use Agents Can Actually Do: a Capability Audit — DutyStation News