Nobody Read This Code: Building AI Factories You Can Actually Trust

Once agents write most of the code, the bottleneck moves. It’s no longer implementation but everything around it: how work gets classified and specified, how agents are orchestrated, how changes get merged, and above all how you verify output that no human has read line by line.

This session goes inside the AI factories we’ve built and run in greenfield and brownfield codebases. We compare the patterns: staged pipelines from issue to spec, implementation, agent review and PR; loop-based setups; and factories driven by a brief, mission or digital twin instead of tickets. We cover harness engineering, merge queues under agent load, CI that wasn’t built for hour-long jobs, and verification that goes past green tests, like agents proving a feature works end to end.

Real architectures, real failure modes, and the things we threw away. For engineers who want to build a factory, not just watch a demo.