- Sources: HumanLayer write-up, HN 49023019
- Summary: A HumanLayer write-up by Dex Horthy, tied to an AI Engineer talk, argues that fully automated coding pipelines with no human review degrade codebases over time even as the models pass benchmarks. The stated failure mode is that agents optimize for tests scored in seconds and carry no penalty for eroding maintainability, whose cost surfaces over weeks. The proposed alternative front-loads four human planning phases (product requirements, system architecture, program design, vertical slices) before agents implement, targeting a 2x to 3x speedup with a human owning the outer review loop rather than a hypothetical 10x to 100x lights-off factory.
- Comments: HN commenters debate whether the framework restates conventional up-front design, and report that review, not code generation, becomes the real bottleneck when many agent loops feed one merge gate.
- Why it matters: It reframes the practical ceiling on coding-agent throughput as review capacity and maintainability, not model quality or harness tooling.
send feedback on this story