Dex Horthy explains why AI coding-agent throughput does not automatically produce maintainable software. He contrasts manual, agent-assisted and lights-out software factories, drawing on HumanLayer's experience to argue that removing human code review can leave teams struggling with failures in codebases they no longer understand.
Dex Horthy examines reinforcement-learning reward signals and coding benchmarks, showing why passing tests can reward functional correctness without measuring longer-term architectural quality. He discusses longer-horizon verifiers, then proposes product review, architecture, program design and vertical implementation slices to reduce rework while retaining human ownership of generated code.
Watch on YouTube
