A passing audit is not delivery: one agent factory's lessons on proving work actually landed
Oct 27, Karell Ste-Marie on Foundry, his Architect, Worker and Auditor agent setup: audited candidates that never reached main, a recovery watchdog that read prose instead of verdicts, and a fix left uninstalled rather than forced past a timed-out approval.
Source: How I Built Foundry, Part 2: A Passing Audit Is Not Delivery
Karell Ste-Marie runs Foundry, a setup where an Architect agent plans, a Worker implements, and a separate Auditor checks. Part 2 is about what broke after the clean diagram, and the short version is that getting agents to write code was the easy part. Proving what had actually changed was not.
The biggest correction: Foundry had audited branches that had never been integrated. It could prove "candidate exists" and "candidate audit passed," but not "the product now contains this work." So every change gained an explicit integration disposition, and the lifecycle became literal: candidate produced, candidate verified, integration pending, integrated on main, milestone closed. An audit pass certifies one exact commit SHA, not a merge. A merged PR doesn't count either until someone reads back the default branch.
Other lessons follow the same pattern. Workers can't close their own work records; a narrow Closure Auditor adapter can close exactly one bound issue once independent evidence exists. Every external write gets read back. When Ste-Marie switched delivery from pull requests to direct fast-forwards on main, the old PR-based closure adapter was retired and set to fail closed instead of guessing. The recovery watchdog first ignored some auditor failures, read summary prose instead of structured verdicts, and mistook old repair cards for live ones.
The ending is the point. The watchdog fix passed an independent review, but a protected policy edit hit a timed-out approval prompt. The fix sits in a draft PR, recorded as not installed, and nothing forced it through another path. "If a system bypasses a protected approval because it is tired of waiting," he writes, "it has proved that governance is decorative."
Why it matters for code review
For anyone running agents around the clock, this is the gap between a green review and a shipped change. Review verdicts are claims about one commit; delivery is a separate claim that needs its own evidence and its own authority. A factory that runs overnight piles up "audited but never landed" work unless that state is tracked explicitly. And an honest "not done" in the morning beats a queue that quietly routed around its own approval gate.