Always-On Agents Hand You the Review Work

OpenAI shipped dots yesterday, and I read the launch post between two runs of my own agents. They are always-on: a cloud computer each and a browser each, working toward your goals 24/7, connected to over 4,000 apps. One early tester’s dot spotted an invoice he had forgotten to send, drafted it, and only sent once he approved.

The workflow detail is the one I keep coming back to. A dot scopes small bugfixes, builds and tests them, then hands back a finished PR with a short video of what changed. Finished work arrives before you’ve thought to ask for any of it.

John Allspaw’s 2012 essay On Being A Senior Engineer says mature engineers beg for the bad news, asking peers what could be missing and how the design will not work, then asking them to punch holes in the thinking. Underneath it is a quieter claim: nothing an engineer makes stays only in their hands, so peer review is where the design actually gets better.

That question used to get asked at design time, over a page of notes, by someone with the context to break it. Now it gets asked at review time, of a diff I didn’t write, at whatever speed I read diffs. Nobody in the loop interrogates the agent’s own output; it hands it over done.

I want both halves of this. The dot would have caught that forgotten invoice just as well, and the pile of PRs waiting on my judgement grows the same week.

An always-on agent asks itself nothing, so begging for the bad news is the last step in the pipeline where a human is still required.