Concourse runs coding agents inside gates you hold. Every run arrives as a proposal — the diff, the passing tests, the evidence behind each acceptance criterion — and nothing merges until it’s earned the right to. Trust becomes a dial you turn, not a default you accept.
Every agent starts fully gated — you confirm what “done” means, approve the plan, review every merge. As the work proves itself, you turn gates off one at a time. The deterministic checks never turn off at all. This is the road we’re building: today, every plan is human-approved.
When a run is ready to merge, Concourse shows you the proof, not a promise: the satisfied acceptance gate, every changed file, the tests that back each criterion. Approve in one click — or open the diff and look closer.
Most tools collapse “the step ran” and “the result is acceptable” into one green check. Concourse keeps them apart. Deterministic facts can hard-stop a run; a model’s judgment can inform and loop it — but never gets the last word on a merge.
Tests, coverage, CI, secret-scan. Deterministic and binary — they either pass or they block the merge. No model in the loop, no negotiation.
A reviewer model scores the work and explains its reasoning. A low score sends the run back to revise — it never silently approves a merge on its own.
Every run moves through the same phases — and stops at any one of them the moment it needs you. Pause it, inspect the evidence, send it back a step, resume where it left off. Nothing is a black box.