Deploy gates

A deploy gate is a check that runs a set of evaluations against a change before it merges and blocks the merge on failure. For agents the change is typically a diff touching prompts, context assembly, tool definitions, or model configuration. The gate evaluates the modified agent and returns its verdict before the change ships.

Each evaluation in the set pairs an input with a judgment about the agent's output. The input is a scenario the agent must handle: a user request, a conversation state, or a situation that calls for a particular tool. The gate runs the modified agent on that input and scores the result by comparing it to a known-good answer, asserting properties of it (the right tool was called, the arguments were valid, a required constraint held), or having a model grade it against a rubric. An evaluation passes when its score clears its threshold.

A gate's checks accumulate the way regression tests do. Each evaluation encodes a failure the team has decided must stay fixed, and once added it runs against every future change, so coverage grows with the agent's history. A red result points at a specific failure mode that has happened before.

A gate reads diffs, so its scope is the regression causes that originate in the team's own changes, the ones that arrive through the merge process. Within that set, every change is checked before it ships.