--- schema: 1 type: video title: Why Software Factories Fail description: >- In July 2025 Dex Horthy turned the lights off: an agent software factory where nobody read the code. It fell apart. An issue appeared that no amount of prompting could fix, the site was down, users were furious, and he was digging through a codebase he had stopped reading three months earlier. His claim is that this is not a skill or scale issue, and no harness or extra tokens fixes it, because it is a model training problem. Coding models are reinforced on one thing, did the test pass without breaking another, and nothing in that reward penalizes bad architecture, whose cost shows up months later. So they get better at passing tests and no better at keeping a codebase maintainable. That is why Claude Code went from nothing to billions while tools with the same read, write, and edit commands did not: it was the first model trained against the harness it ships in. But maintainability is far harder to verify than a green test, and as Horthy puts it, if a model knew what good code looked like it would already write it. So for now you are stuck reading the code, which is fine, because you can still move fast. His fix is to turn the lights back on and plan up front: product review, system architecture, the underrated step of program design down to types and call graphs, then vertical slices. Thirty minutes of alignment saves hours of review, and a good PR becomes a joy to read instead of slop to untangle. Speaker info: - https://x.com/dexhorthy - https://linkedin.com/in/dexterihorthy - https://github.com/humanlayer/12-factor-agents Timestamps: 0:00 - The narrative: you are the bottleneck, just ship more 1:28 - The cracks: outages and falling PR review quality 2:20 - The thesis: the harness is not enough 3:36 - A brief history of the software factory 5:52 - The agentic factory and turning the lights off 7:30 - Why it fails: the July 2025 lights-off experiment 8:56 - Models cannot maintain codebase quality 10:12 - Why Claude Code won and how coding models are trained 13:18 - Verifying maintainability and better benchmarks 14:58 - Turning the lights back on: plan up front 17:16 - Too many bad PRs, and closing advice --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/Ib5GBkD555M-harness-engineering-is-not-enough-why-software-factories-fail Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/Ib5GBkD555M' markdown: 'https://liquidgold.cc/watch/Ib5GBkD555M.md' llms: 'https://liquidgold.cc/watch/Ib5GBkD555M/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T06:45:24.045Z' language: null children: [] videoId: Ib5GBkD555M durationSeconds: 1158 publishedAt: '2026-07-23T16:30:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 36 endSeconds: 90 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=36' heldReason: rights - ordinal: 1 text: null startSeconds: 90 endSeconds: 101 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=90' heldReason: rights - ordinal: 2 text: null startSeconds: 101 endSeconds: 145 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=101' heldReason: rights - ordinal: 3 text: null startSeconds: 101 endSeconds: 145 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=101' heldReason: rights - ordinal: 4 text: null startSeconds: 157 endSeconds: 222 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=157' heldReason: rights - ordinal: 5 text: null startSeconds: 157 endSeconds: 222 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=157' heldReason: rights - ordinal: 6 text: null startSeconds: 222 endSeconds: 232 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=222' heldReason: rights - ordinal: 7 text: null startSeconds: 232 endSeconds: 353 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=232' heldReason: rights - ordinal: 8 text: null startSeconds: 232 endSeconds: 353 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=232' heldReason: rights - ordinal: 9 text: null startSeconds: 364 endSeconds: 452 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=364' heldReason: rights - ordinal: 10 text: null startSeconds: 364 endSeconds: 452 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=364' heldReason: rights - ordinal: 11 text: null startSeconds: 463 endSeconds: 538 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=463' heldReason: rights - ordinal: 12 text: null startSeconds: 463 endSeconds: 538 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=463' heldReason: rights - ordinal: 13 text: null startSeconds: 463 endSeconds: 538 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=463' heldReason: rights - ordinal: 14 text: null startSeconds: 538 endSeconds: 548 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=538' heldReason: rights - ordinal: 15 text: null startSeconds: 548 endSeconds: 621 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=548' heldReason: rights - ordinal: 16 text: null startSeconds: 632 endSeconds: 805 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=632' heldReason: rights - ordinal: 17 text: null startSeconds: 632 endSeconds: 805 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=632' heldReason: rights - ordinal: 18 text: null startSeconds: 805 endSeconds: 816 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=805' heldReason: rights - ordinal: 19 text: null startSeconds: 816 endSeconds: 903 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=816' heldReason: rights - ordinal: 20 text: null startSeconds: 816 endSeconds: 903 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=816' heldReason: rights - ordinal: 21 text: null startSeconds: 913 endSeconds: 1043 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=913' heldReason: rights - ordinal: 22 text: null startSeconds: 913 endSeconds: 1043 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=913' heldReason: rights - ordinal: 23 text: null startSeconds: 913 endSeconds: 1043 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=913' heldReason: rights - ordinal: 24 text: null startSeconds: 913 endSeconds: 1043 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=913' heldReason: rights - ordinal: 25 text: null startSeconds: 1043 endSeconds: 1054 watch: 'https://liquidgold.cc/watch/Ib5GBkD555M?t=1043' heldReason: rights --- # Why Software Factories Fail In July 2025 Dex Horthy turned the lights off: an agent software factory where nobody read the code. It fell apart. An issue appeared that no amount of prompting could fix, the site was down, users were furious, and he was digging through a codebase he had stopped reading three months earlier. His claim is that this is not a skill or scale issue, and no harness or extra tokens fixes it, because it is a model training problem. Coding models are reinforced on one thing, did the test pass without breaking another, and nothing in that reward penalizes bad architecture, whose cost shows up months later. So they get better at passing tests and no better at keeping a codebase maintainable. That is why Claude Code went from nothing to billions while tools with the same read, write, and edit commands did not: it was the first model trained against the harness it ships in. But maintainability is far harder to verify than a green test, and as Horthy puts it, if a model knew what good code looked like it would already write it. So for now you are stuck reading the code, which is fine, because you can still move fast. His fix is to turn the lights back on and plan up front: product review, system architecture, the underrated step of program design down to types and call graphs, then vertical slices. Thirty minutes of alignment saves hours of review, and a good PR becomes a joy to read instead of slop to untangle. Speaker info: - https://x.com/dexhorthy - https://linkedin.com/in/dexterihorthy - https://github.com/humanlayer/12-factor-agents Timestamps: 0:00 - The narrative: you are the bottleneck, just ship more 1:28 - The cracks: outages and falling PR review quality 2:20 - The thesis: the harness is not enough 3:36 - A brief history of the software factory 5:52 - The agentic factory and turning the lights off 7:30 - Why it fails: the July 2025 lights-off experiment 8:56 - Models cannot maintain codebase quality 10:12 - Why Claude Code won and how coding models are trained 13:18 - Verifying maintainability and better benchmarks 14:58 - Turning the lights back on: plan up front 17:16 - Too many bad PRs, and closing advice --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/Ib5GBkD555M-harness-engineering-is-not-enough-why-software-factories-fail Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/Ib5GBkD555M · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.