Five levels of oversight for AI coding agents
can be divided into five levels based on how much stands between their work and a live service. At Level 1, Assist, a person writes and owns every line while the agent mainly speeds up typing. At Level 2, Review, the agent creates complete changes, but a person checks each one before release. This level suits work involving money, security, or customer data.
At Level 3, , the agent creates and tests changes, while automatic checks and a person stop anything that fails. This is presented as the best fit for most real product work that must move quickly. At Level 4, Orchestrate, several agents work at the same time under an experienced person who directs and reviews them, and the agents cannot publish their own changes. At Level 5, , agents work unattended, such as overnight, and a person reviews the batch later.
This last level is suitable for disposable experiments, not a live product that a business relies on. A higher level means less supervision, not better work.
Key points
- Level 1 keeps all coding and responsibility with the person.
- Level 2 lets the agent make full changes but requires every time.
- Level 3 combines agent-written and agent-tested changes with automatic and human checks.
- Level 4 runs several agents in parallel under an experienced person who controls publication.
- Level 5 runs without supervision and is recommended only for work that can be discarded.