A quiet but significant permission change
On September 2, 2026, Anthropic rolled out background computer use for Claude Cowork and Claude Code on macOS. Until now, letting an agent operate your machine meant handing over the mouse and keyboard entirely — the agent took the foreground, and you waited. The new mode lets an agent click, type, and switch between applications in the background, while your own foreground work continues untouched.
The feature is opt-in, limited to Pro and Max subscribers, and ships switched off by default. When an agent needs the entire display — for a modal dialog it cannot dismiss any other way, say — it asks first. If you step away from your machine mid-task, the agent keeps working rather than pausing and waiting for you to come back.
It's a small UX change on paper. In practice, it removes the single biggest friction point in using computer-use agents for real work: the fact that using one meant you couldn't use your own computer.
Why "foreground-only" was the real bottleneck
Computer-use agents have existed for a while, but the constraint that made them impractical for daily work was rarely the model's competence — it was the exclusivity. A QA pass through a staging UI, a repetitive data-entry task in an admin console, or a scripted click-through of a third-party dashboard without an API: these are exactly the jobs an agent is good at, and exactly the jobs nobody wants to babysit for twenty minutes.
Background execution changes the cost equation. A task that previously required blocking your own machine for its duration now runs alongside your work. That turns "occasionally worth automating" into "worth automating by default," for a wider set of tasks.
Where this fits into a real workflow
For the kind of client engagements webhani runs — building and maintaining internal tools, integrating SaaS admin panels that lack decent APIs, running UI regression passes before a release — background compute changes what we'd reach for an agent to do at all. A few concrete shapes:
- Regression walk-throughs. Point an agent at a staging environment to click through a critical user flow (signup, checkout, settings changes) after a deploy, while the engineer who shipped the change keeps working on the next ticket.
- Admin console housekeeping. Bulk operations in a vendor dashboard that only exposes a UI — bucket permissions, user role cleanup, report exports — handed to an agent as a background task instead of blocking a person's afternoon.
- Exploratory UI testing. Running a battery of "what happens if a user does this" scenarios against a new feature branch, in parallel with the next feature's implementation.
None of these are new categories of task. What's new is that they no longer compete for the same screen the person doing the "real" work is using.
The governance question doesn't go away
Background execution does not remove the need for oversight — it just changes its shape. A few things worth deciding before turning this on for a team, rather than after:
- Scope the environments. Background agents should generally point at staging or sandboxed accounts, not production consoles with billing access, unless the task and its blast radius are well understood.
- Keep a record. Whatever the agent did while you weren't watching needs to be reviewable after the fact — screenshots, action logs, or a summary at minimum. "It ran while I was in a meeting" is not an audit trail.
- Treat the full-screen prompt as a real gate, not a nuisance. The one moment the agent asks for your attention is precisely the moment something outside its normal operating envelope happened. Training a team to reflexively approve that prompt defeats the point of it.
- Decide who's accountable for agent actions in shared accounts. If an agent files a support ticket, changes a permission, or submits a form on a company's behalf, someone needs to own that action the same way they'd own a teammate's.
Our take
The interesting part of this release isn't the technical mechanism — it's that Anthropic shipped the permission boundary as carefully as the capability itself: off by default, scoped to a specific OS and subscription tier, and gated behind an explicit ask for full-screen control. That's the right order of operations for a feature that runs unattended.
For teams evaluating whether to use this, our advice is the same as with any agent capability that expands unattended: don't skip the step of writing down what "acceptable to run without a human watching" means for your environment before you rely on it in one. Background compute makes agents genuinely more useful for real operational work. It also means the review discipline has to keep up with the convenience.
Reference: AI News Sept 4: GPT-6 Astra, AI Outage, Claude Background