We built Agent Role Contracts after repeatedly seeing a gap between what an AI agent was told to do and what the surrounding workflow had actually declared it was allowed to do. The package checks those declarations before handoff. The starter shows a task inside scope, an expected failure outside the declared write scope, and the repaired task passing again. It is deliberately not a sandbox or execution authority. A PASS only means the supplied declarations are internally consistent. The first public version checks role references, declared authority, explicit task routing, write-scope containment, reviewer separation, required inputs, and handoff consistency. The starter needs no API key or agent runtime after cloning, so the behavior is easy to inspect directly. I’m most interested in feedback on the boundary: should declaration-level preflight remain a small standalone primitive, or should these checks be absorbed into agent runtimes and orchestration systems?

Agent Role Contracts

Catch AI-agent scope conflicts before handoff

Catch AI-agent scope conflicts before handoff