HOW WE BUILD
We run our own fleet of AI coding agents. Each one works a ticket in its own isolated environment and opens a pull request. Models from different families review that change independently, automated tests run against it, and an engineer who is accountable for the work makes the call.
THE FLEET, AS OF SEPTEMBER 24, 2026
Read from our orchestrator, the system that dispatches the agents and tracks every ticket and pull request they touch.
THE PIPELINE
More detail on each gate is on how we guarantee quality.
A REAL REVIEW
From a change to our own orchestrator, July 2026, quoted from the pull request and lightly shortened.
Review · two models, split verdict
Split verdict. Claude Code (worktree) approves; Codex (diff-only) requests changes.
Claude Code · approve
Ran the type checks and tests, confirmed the safety check sat in the right place, and raised two non-blocking issues, including one where a typo in configuration would have switched a feature on instead of off.
Codex · request changes
Found that the code did not do what the pull request said it did: names the description promised would be rejected were in fact accepted, and no test covered the case. Suggested a stricter check.
The engineer's call
Codex was right about the facts and wrong about the fix. The stricter check would have failed open: anything it didn't recognise would have been treated as trusted. The engineer overruled the suggested fix, kept the safe behaviour as the documented contract, added the missing test, corrected the description, and fixed both of Claude Code's issues too.
Matching broadly fails safe … while [the suggested check] would fail open.
Agreement between models is cheap. What you're paying for is a process that notices when they disagree, and a person who is accountable for the decision.
The same fleet and the same gates work on your codebase, under your review.
