The unlock for agents will not come from trusting the model. It will come from building a system that understands what “done” means and improves its ability to verify it. Imagine asking an agent to write release notes across 1,247 commits. It returns a polished draft, but the comparison API paginated at 1,000 commits and the agent never followed the cursor. The result looks complete even though 247 commits never entered the context. These checks do not prove that the release notes are well written. They establish something more fundamental: that the agent received the complete inputs and produced an internally consistent result. Verification produces the fact, and authorization consumes it. Do not promote the model. Promote the loop, and let the loop learn. Illustrative code: Blog: