Ephesus — notes from the build

Ephesus — notes from the build https://ephesushq.com/ Notes from building Ephesus, a multi-agent harness you govern as its architect. Everything here comes out of the project's own records. en Fri, 04 Sep 2026 00:00:00 GMT Why the crew works in its own checkout https://ephesushq.com/blog/their-own-checkout https://ephesushq.com/blog/their-own-checkout Fri, 04 Sep 2026 00:00:00 GMT design Running several agents in the architect’s working tree is fine until it isn’t. The fix was a worktree each — and the interesting part was what that broke. Forty green tests, nine killed mutants, and three ways past the gate https://ephesushq.com/blog/forty-green-tests-three-bypasses https://ephesushq.com/blog/forty-green-tests-three-bypasses Thu, 03 Sep 2026 00:00:00 GMT method A new coverage gate passed its own tests, survived mutation testing, and was defeated three times in ten minutes by someone actively trying to walk through it. The recurring defect of this codebase is a check that cannot fail https://ephesushq.com/blog/check-that-cannot-fail https://ephesushq.com/blog/check-that-cannot-fail Wed, 02 Sep 2026 00:00:00 GMT method Five instances found in a single day, in five different subsystems, by five different routes. The pattern was not carelessness — every one of them was written by someone trying to be careful. The MVP ships one engine, and says so https://ephesushq.com/blog/the-mvp-ships-one-engine https://ephesushq.com/blog/the-mvp-ships-one-engine Wed, 02 Sep 2026 00:00:00 GMT decision Ephesus is designed around an engine-agnostic adapter seam. It ships with exactly one engine wired — and the decision record exists mostly to stop that being misread in both directions. A green suite over 1,406 lines nothing could reach https://ephesushq.com/blog/lines-nothing-could-reach https://ephesushq.com/blog/lines-nothing-could-reach Mon, 31 Aug 2026 00:00:00 GMT defect A milestone shipped a subsystem that was fully implemented, fully tested, and never once called by the application. Every test passed. The feature did not exist. A research department that must cite a commit or die https://ephesushq.com/blog/cite-a-commit-or-die https://ephesushq.com/blog/cite-a-commit-or-die Sat, 29 Aug 2026 00:00:00 GMT design The Stoa reads other people’s repositories and reports what it finds. The rule that makes it trustworthy also caught it citing a commit that existed in no repository at all. A project whose primary mission is its own improvement https://ephesushq.com/blog/a-mission-that-is-itself https://ephesushq.com/blog/a-mission-that-is-itself Fri, 28 Aug 2026 00:00:00 GMT design The Gymnasium lets the company propose changes to itself. Everything interesting about it is in the constraints — a falsifiable metric before any human reads it, and a ledger that keeps the rejections.