Works with every agent, trusts none of them
Arcus plugs natively into Claude Code, Codex, Copilot, and Cursor, reading exactly what each one shipped, and checks it independently of the agent that wrote it.
Reads what each agent actually shipped
Arcus reads the diff, the commit, and the PR context in the shape each agent produces it, so the check is grounded in what really happened, not a lowest-common-denominator summary.
Reads the PR and commit trail Claude Code leaves behind to check the diff against the ticket it was given.
Follows Codex’s run history and generated tests to see what was covered and what was assumed.
Checks Copilot’s inline completions against the surrounding intent, not just whether the code compiles.
Reads Cursor’s session history to see what changed and why, then verifies it independently.
One score, whichever agents your team runs
Most teams aren’t running one agent, they’re running three or four, often on the same codebase in the same week. Arcus doesn’t care which one touched a given PR, the release confidence score is calculated the same way every time.
No agent grades its own work
An agent checking its own output has no reason to find the gap it just introduced. No amount of prompting fixes that, so verification has to come from somewhere the agent doesn’t control, a separate process, not a feature bolted onto the one that wrote the code.
Arcus doesn’t share context, memory, or incentives with the agent that generated the code. It checks behaviour against intent, not code against its own explanation of itself.
Claude Code, Codex, Copilot, or Cursor, the release confidence score means the same thing regardless of which agent your team happened to run that day.
Context from the rest of the stack, too
Which agent wrote the code is only half the picture. Arcus also pulls in the ticket, the design, and the history around it, so the check is grounded in intent, not just in diffs.