Does computer use replace API integrations?

No. It answers for systems you cannot integrate, either because no programmable interface exists or because opening one would take longer than the project lasts. When an API covers the operation, it remains the better option: it is easier to test, replay and log.

Is a computer use agent reliable in production?

Its reliability depends on the interface it drives. A screen that changes layout, an unexpected modal window or an unusual loading time is enough to derail the sequence. In production this mode calls for active monitoring and a human approval point on every write operation.

What does the EU AI Act say about an agent driving an interface?

The text draws no distinction between technical modes. Since 2 August 2026, the Article 50 transparency duties and the penalty regime apply. The omnibus package, adopted in June 2026 and in force since 27 July, pushed the substantive requirements for Annex III high-risk systems, including detailed documentation and reinforced human oversight, back to 2 December 2027.

How do you trace the actions of an agent that clicks?

Instrumentation belongs on the layer that drives the browser or the workstation, rather than on the model itself. The log has to keep the requested task, the action sequence, the identity used and the observed result. Without that layer, audit comes down to screenshots, which fails in regulated sectors.

Should you wait for the technology to mature before deciding?

Waiting costs nothing on high-stakes use cases, where integration remains the right answer anyway. On tasks shelved for lack of an API, an instrumented trial on a narrow perimeter gives you an internal measurement, more useful in committee than any vendor benchmark.

Computer use or API: deciding AI agent architecture

What is a computer use AI agent?

OpenAI shipped GPT-6 Astra on 3 September 2026 and made computer use its headline claim: the model drives software through its interface, the way a person would, instead of going through an application integration. Two days earlier, Anthropic had released Claude Fable 5.1, which also gets better at computer use without building its pitch around it. The capability is settled on both sides. What architecture boards are working through this autumn is which mode they take as the default, and what they accept losing by taking it.

Computer use is a model’s ability to operate software through its graphical interface. It reads the screen, moves the cursor, fills in fields and clicks. The target application needs no API. The agent works on the rendered view, in the same place a human operator works.

The capability itself is not new. Anthropic opened it in preview back in late 2024 and vendors have refined it since. What changed in September 2026 is its place in the sales narrative: it moved from side feature to architectural promise.

Why is this decision reaching architecture boards now?

Because the promise speaks first to the budget holders. An agent that drives the screen can reach applications that expose nothing: a thick client from the 2000s, a vendor-locked business package, a supplier extranet with no programmable interface. Most large European enterprises run a few of those. They stall automation projects at the estimation stage far more often than any limit of the models does.

Since Astra shipped, the same line keeps coming back to us from engineering leaders: integration is optional now. Taken at face value, the line is wrong. It also holds enough truth to deserve a written decision rather than a corridor debate, because the two modes are not governed the same way.

Computer use or API integration: how do you decide?

Use the API whenever one exists and covers the operation. Keep computer use for systems you cannot integrate, or cannot integrate within the time the project has. The first mode can be tested and replayed identically. The second has to be watched.

Four questions settle it, use case by use case.

Does an API cover the operation end to end, or only the read side?
What does an error cost: a line to fix tomorrow, or an irreversible write in a production database?
How will you know the agent got it wrong, and after how long?
Who approves before the write, and at which point in the chain?

The fourth question is the one that stays unanswered in the scoping workshops we run. The agent is scoped technically, its write perimeter is documented, and nobody can say who picks up the alert when it lands at seven on a Friday evening.

Why do benchmarks not decide for you?

Published scores on OSWorld, the task suite that became the reference for computer use, move a lot depending on the evaluation protocol. The same model posts markedly different scores between a standard configuration and a reinforced harness. Each vendor publishes the setup that flatters it, which is fair enough. The gap stops being a methodological footnote the day it is used to justify an architecture choice in a steering committee.

An agent’s success rate on a public bench says little about its behaviour on your ERP, with your permissions and your customised screens, on a month-end closing day. The only measurement that commits you is the one you produce on your own tasks, with your own data. Gartner expects 40 % of agentic projects to be cancelled by 2027. We keep seeing the same cause at the start: an architecture decision settled on figures nobody replayed in house.

What does this change for a European enterprise?

Traceability first. An agent that clicks leaves a thin trail: a series of screenshots and coordinates, with no transaction identifier and no return code. An agent that calls an API leaves an auditable record, with the call and its timestamped response. Since 2 August 2026, the Article 50 transparency duties and the EU AI Act penalty regime apply. Several member states, France and Belgium among them, have yet to formally designate their competent authorities. The omnibus package, adopted in June 2026 and in force since 27 July, also pushed the substantive requirements for Annex III high-risk systems back to 2 December 2027. That buys time on heavy documentation and none at all on being able to answer a regulator asking who did what, when and on which system.

Identity second. An agent driving a screen inherits a session, and with it the permissions of whoever opened it. Many organisations discover at that point that they have no service account for an agent, and no clear answer on accountability for a write performed under the account of a colleague on leave. Deloitte measured in 2025 that only 21 % of organisations deploying agents had mature governance. Computer use widens the surface to govern before that figure has moved.

Where should you start?

With an inventory. The pilot comes after. List the tasks you wanted to automate and shelved because no API was available. We keep finding that list in IT departments, in the form of tickets closed as out of scope. It marks the real ground for computer use, narrower than the sales promise and wider than nothing.

Then take a single task from that list, one with a low error cost, and instrument it before automating it: an action log and a named human approval point. S&P Global reported in 2025 that 46 % of organisations abandoned their prototypes before production. The ones that get past that point tend to have settled observability before writing the first prompt.

Our AI maturity assessment covers this architectural dimension in three minutes, alongside governance and industrialisation, using the Koneetiv framework (2026 edition). For the wider picture, see AI agents in the enterprise, the LOOP™ AI governance methodology and our analysis of moving from pilot to production.

Computer use or API: how to decide your AI agent architecture