Why
Tracking where computer-use agents actually stand in production, not in demos. The a16z piece is built on interviews with real operators — a CPG data platform running 15-20M portal interactions a month, a systems integrator running 27 live workflows — not just benchmark numbers. A useful checkpoint for judging whether an "agent operates a browser or UI" idea is worth building now versus still a research toy.
The headline number is real and the argument is that it stopped mattering. OSWorld-Verified, a real-desktop benchmark, went from 42% a year ago to 85% today, above the roughly 72% human baseline. But once models cleared a good enough bar, buyers stopped comparing models and started buying infrastructure — verification, escalation, error handling. The model is no longer the product.
And the other end of the same question is a much smaller, much cheaper answer that this catalogue kept as a separate card for no good reason. Alibaba's page-agent is an in-page GUI agent: one script tag or npm install, and natural language operates that page's UI. No screenshots, no browser extension, no headless browser — it reads a text-based DOM, connects to any LLM you bring, has a beta MCP server, and is built on browser-use. The general capability question and the specific embedding decision are the same question at two scales, and holding them together is what makes either useful: the a16z piece says the bottleneck is now operations, and page-agent is the case where almost all of the operations disappear because the agent only ever touches one page you control.
The failure mode that matters most is the one neither half advertises. It is not the UI click failing — it is silent failure with no verifiable signal at run time. An insurance claim that "succeeded" on screen and stalls two days later on a phone call nobody logged. That is why verification, not accuracy, is what buyers pay for, and it is the same shape as the-harness-not-the-model: the interesting engineering is around the model, not in it.
How it works
Where the benchmark went, and why it stopped mattering
| A year ago | Today | |
|---|---|---|
| OSWorld-Verified | 42% | 85% |
| Human baseline | ~72% | ~72% |
| What buyers compared | Models | Infrastructure |
Once models cleared the bar, the purchase became verification, escalation and error handling. A common pattern: an agent runs a workflow once, caches it as deterministic code, and is only re-invoked when something breaks — so cost falls over a workflow's lifetime rather than staying flat.
The cost comparison that decides whether to build
| Option | Rate | Against agent inference |
|---|---|---|
| Agent inference | ~$6-8/hr ($3-15 range) | — |
| Offshore BPO | ~$10/hr | Roughly break-even today |
| Fully-loaded US labour | $30-45/hr | 70-80% margin |
Two scales of the same capability
| General computer use | In-page agent (page-agent) |
|
|---|---|---|
| Surface | Any application | One page you own |
| Input | Screenshots, accessibility trees | Text-based DOM |
| Needs | Extension or headless browser | A script tag |
| Operations burden | Verification, escalation, error handling | Almost none — the surface is yours |
| Fails at | Silent failures downstream | Anything the DOM does not express |
| Model | Whatever the vendor ships | BYO, any LLM; beta MCP server |
The right way to read the pair: the general case is where the operations layer is the product, and the embedded case is the one where you can skip most of it — because you control the surface, so a failure is visible in your own application rather than two days later on somebody else's phone call. A landing-page or SaaS copilot, or smart form-filling, is the natural first use.
The failure mode to design against
Not the click that fails — the one that appears to succeed. No verifiable signal exists at run time, so the workflow reports done and the real state diverges silently. Verification is not a feature bolted on afterwards; it is the reason the infrastructure layer exists, and it is why the honest first question for any browser-agent idea is how would I know this went wrong rather than can the model do it.
The authoring end of the same question
This card asks whether an agent can drive a computer and concludes that once it can, what you buy is operations. demonstration-is-not-a-specification arrives at the same conclusion from the other end — how a task gets stated — by way of the screen-recording interfaces that a startup and a large vendor shipped weeks apart in 2026. The short version: recording lowers the cost of authoring a task and leaves the cost of running it reliably exactly where it was.
← All Dev Notes · Workspace Index · Top ↑ · Open on jaylabs.xyz →