Send a cloud agent. Get back a mergeable pull request.

The run streams into the app as it happens — the commands, the files it touched, the reasoning. That matters the first few times, when you are deciding whether to trust it, and it matters again when something goes strangely.

After that, the point is not to watch. The merge queue and keep-mergeable exist so that the common cases never need you at all: something breaks, a fix dispatches, the pull request lands, and the first you hear of it is that it merged.

Fixing a pull request is one of three things you can send. A review response reads the conversation and addresses what reviewers actually asked for. And a freeform task takes a prompt and a repository and opens a pull request of its own.

Some things an agent should not decide. A merge gate that your repository's policy says a human must clear. A credential the sandbox does not have. A product question with two defensible answers.

A run that stops for one of those reasons is recorded as needing a human, with the agent's own explanation, and it is deliberately not recorded as a failure. The difference matters: a failure is indistinguishable from a crash, and both of Talyn's retry loops will stand down on a stated refusal rather than spending three more runs discovering the same thing.

On Talyn Fleet, each task gets its own Firecracker microVM on our hardware, running on your Claude or ChatGPT subscription. The credentials are attached by a proxy outside the machine, so no token is ever inside the box running your code, and the machine is destroyed when the task ends. With PostHog Code the run happens in their cloud under your own account instead.