Coding agents made writing code almost free. Maintaining it did not get any cheaper. Once most of your codebase is written by agents, the bottleneck moves from generation to everything after: knowing what the system actually does, keeping it stable, and deciding what deserves to survive.

This talk is about that second phase, from a team running large production AI systems where most of the code, and the prose instructions the agents run on, is written by coding agents. We start with the failure modes that only show up at scale and only after a few months:

Green but wrong. Tests pass, the agent says done, and production still breaks on the path nobody specified.
Invented work. Whatever sits in the model’s context becomes a reason to act, so it builds checks nobody asked for.
Duplicated truth. The prose says one limit, the code says another, and nothing throws an error.
Sprawl. Branches and variants multiply faster than any human can review, and the dead ones still look alive.

Then the patterns that actually help:

Replays of real tasks, and evals written from expert corrections, as the evidence that counts, instead of line-by-line review
Evaluation owned by domain experts, so “it works” means the business agrees, not just CI
Guardrails in the runtime and drift detection in the prose that agents cannot quietly route around
A second agent that checks with no shared context, and weekly pruning with evidence instead of the Thanos snap
What the human owner must keep in their head, and what can safely be delegated

You will leave with a checklist of eighteen questions for auditing your own AI-coded repo before it starts auditing you.