AI coding with human approval
A workflow where the agent does the work and people stay in charge.
The problem
AI agents that change code on their own are fast, but teams can't see what will change or stop a mistake before it lands.
The workflow
- A developer, product owner or founder describes the change from the web or a phone.
- The agent reads only the allowed project and proposes a plan with files, risky steps and cost.
- A person approves, changes or cancels it.
- The change is made on the developer's own computer behind the danger wall, tests run, and the result is reported back.
- Anyone can review the files; one click undoes everything.
Why it works
- Nothing happens without a person's OK, and risky steps ask again at the moment they happen.
- Done means tested: the 3-try loop fixes failing tests and stops before wasting credits.
- Every action is logged for the team.