Orbital Witness reports merged PRs per engineer roughly doubled after a harness rebuild, and marks three of twelve theses as half-built

Thompson scores his team's harness against the twelve theses in Ryan Lopopolo's harness-engineering repository, the follow-up to OpenAI's February post on shipping with Codex. He reports that on a legacy codebase with the harness switched on, weekly merged PRs per contributor rose from about 3 to the mid-to-high single digits, mostly 6 to 7. On a newer codebase built fully around the harness he reports a swing between the low teens and low twenties a week, about three times the legacy repo's rate. He says every PR clears the same review, type-check and CI gates as before, with agents handling most review and humans checking the rest. He gives these as ranges, and merged PRs are a proxy for output, not a measure of it.

The harness pieces he describes are specific: an isolated Docker stack per worktree so agents run the real system, a nightly bug hunt in which sceptic agents try to refute each finding before a fix is attempted, forbidden paths re-checked on every push to an agent-authored PR, human-only merges, and a git hook that stamps each agent-assisted commit with token and dollar cost. He marks three theses as only partly built: turning feedback into infrastructure, retiring nightly loops that have no end condition, and measuring effectiveness. On the last he writes that "we measure cost, not attention."

Why it matters: The gaps are the useful part. A team running a mature harness says it can price a commit in tokens but cannot yet measure the steering and review time a commit costs a human, which is the metric the playbook says matters. Treat the PR figures as one team's self-report, and the unbuilt list as a checklist for your own harness.