Principles
Keep people in the decision loop
Agents may act on their own within agreed limits, and a person makes the decisions that go beyond them.
Some decisions are cheap to reverse and some are not. An agent can rename a variable without asking. It should not change a production database schema, spend money or alter access rights without a person’s explicit approval.
Rationale
Clear limits let agents move fast on routine work while people stay in control of decisions with lasting consequences. Without limits, teams either approve everything by habit or slow every change to a crawl.
Setting the limits
Write the limits down where both people and agents can read them. Typical limits cover:
- Changes to infrastructure, security settings or data retention.
- Anything that costs money or contacts a customer.
- Deleting data, history or other people’s work.
Review the limits as the team learns. A limit that never triggers may be too loose, and one that triggers on every change trains people to approve without reading.