Guardrails
Guardrails that hold where work lands
Advice is easy to ignore. Headron controls are declared in the repo contract and enforced when a change tries to land, so an agent cannot talk its way past them.
How it works
- Narrow tools
- An agent gets only the tools the task needs, nothing more.
- Sandbox
- Work runs contained, so a bad change cannot reach beyond its boundary.
- Budget
- Cost and step limits are set per task, and the work stops when they run out.
- Approval
- Risky changes wait for a human yes before they land.
- Effects
- What a change is allowed to touch is declared in advance, and checked.
- Audit
- Every action and verdict is recorded as a receipt that outlives the session.
Start
Declare the controls once in the contract and every connected agent works under them, locally or through the Mind API.
Read the docs · Mind API reference · Create an account · Star on GitHub