the day-to-day is less “prompt a chatbot” and more “run a small team.” issues get picked up in parallel worktrees, each agent working in isolation, opening PRs that other agents review before a human ever looks.
the human stays in the loop
direction, taste, and the hard calls stay with the person. the agents advise; they do not decide. approving a scoped credential, merging a PR, choosing what not to build — those are human moves, on purpose.
humans decide; agents advise.
least-privilege is the whole game. every agent gets a scoped, revocable endpoint — switchboard is basically this idea turned into a product. revoke equals kill the endpoint.
it is early, and the tools are opinionated because the workflow is. that is the fun part.