Computer Use Is the Bridge, Not the Moat
An agent clicking buttons feels like the future. The real primitive is whether the system knows what the button means — who may press it, and what breaks if it's wrong.
When an agent opens a browser, moves through tabs, and clicks save, it feels like a line has been crossed. It has — but the visible work is distracting us from the platform shift underneath. The future isn't an AI that gets very good at clicking buttons for you. The real fight is over who defines what the button means.
Three layers: access, meaning, authority
Computer use gives an agent access — it's the universal adapter for a world of software built for humans, and it's genuinely necessary. But a screenshot shows what's on screen without revealing the structure underneath, and a click doesn't tell the agent whether the button is reversible, financially material, or dangerous. Moving a calendar invite looks like changing a time; in the work, it may notify five people and break a commitment. The agent needs the layer above access: meaning.
Trust is not a switch
"Trusted write access" is far too small a way to picture this. An agent might be trusted to read but not write, draft but not send, stage but not deploy, recommend but not approve, touch a sandbox but not production. Every one of those distinctions is semantic. If the system can't tell staging from production, it shouldn't be near the deploy button — and real production systems have been deleted precisely because it couldn't.
Why coding agents came first
It's tempting to say code works because language models are good at text. The deeper reason is that a codebase already has rich work semantics — tests, types, dependencies, version history — so the environment gives the agent feedback without a human answering "is this right?" every thirty seconds. Most knowledge work lacks that density. A calendar's importance hides behind politics; a sales process depends on unwritten history.
The button is no longer the primitive; the action behind it is — described, permissioned, reviewable, reversible. So don't ask only whether an agent can act. Ask whether the product knows what the action means.