Cortad lets a coding agent ask how a change to an AI app will land with real users before that change is merged. When Claude Code, Codex, or Cursor edits a prompt, tool, or flow, it can query Cortad over MCP for who uses that part of the product, what they are trying to do, and how it currently behaves in production. Cortad then runs simulated users against the app on the developer's machine, through the team's own model key, and returns the failing conversations with the file and line to fix. The checks stay on a server the agent cannot rewrite. Its tools can start a run, replay a failure, and read results, and the server has no route for writing a case, a check, or a score. About one conversation in five is held back, picked by a hash of the conversation id, so a later verify can say whether the fix held on inputs the agent never saw. Cortad Intelligence builds that user mix from production conversations, traces, and session replays, which is the context this release puts in the agent's session at edit time. Teams connect a local app with one command and can start free. The release is aimed at a workflow coding agents made routine: the same agent that writes the change also writes the eval, then grades itself on cases that share its blind spots. Cofounder Alsufi is shipping the production read inside that loop, so the next edit can follow how people actually use the feature.