Check your inbox for a confirmation email where you can adjust your preferences and even join additional groups. Async agents are only useful if you can trust what they hand back. In a distributed system, that trust comes down to the runtime you can give them: in the inner loop, before the PR. Ido Pesok at Cognition recently published a post about a milestone worth sitting with. For the first time, his team triggers more Devins asynchronously, from events, schedules, automations, and other agents, than they do in interactive sessions. The agents are no longer something a developer drives. They run on their own and hand back work.
When a developer drove every agent, that developer was also the verifier. They read the diff, ran it against the real system, and decided if it was right. Take the developer out of each step, and you remove the verifier, too. Now the agent has to verify its own work at the speed and volume at which agents generate code. As Ido puts it, async agents are only useful if developers can trust what they return. "An async agent that cannot verify itself is not saving anyone time. It is opening a PR and asking something downstream to grade it. " That reframes the whole problem. Generation stopped being the constraint a while ago. The constraint is verification.
And an async agent that cannot verify itself is not saving anyone time. It is opening a PR and asking something downstream to grade it. A harnessed agent does real work on its own. It writes the code, runs the unit tests it can run locally, exercises its mocks, and reports back green. The trouble is what green means. The agent wrote those mocks. It wrote them to match its own model of how the dependency behaves, and that model is exactly what might be wrong. So the agent tests its change against