Fighting Code Slop: The State of Software Factories - Dexter Horthy

Dex will be speaking at AGNTCon NA on October 22 & 23 in San Jose. Join 3500 other people at the event: https://hubs.la/Q04gbXLb0 Dex Horthy, CEO and co-founder of HumanLayer, argues that engineering teams shipping AI-written code still have to read it. This talk is for engineering leaders and developers building agentic coding pipelines who want to know where the lights-off software factory actually breaks down. His thesis: unattended models will not improve or maintain your codebase quality over time. Models got dramatically better in 2026 at solving problems. They did not get much better at architecture, abstraction, or centralization, and the cost of bad architecture is measured in months and years, not minutes. LINKS AND RESOURCES - Every resource Dex has published on coding agents, chronological: https://humanlayer.dev/resources - HumanLayer: https://www.humanlayer.dev/ - 12-Factor Agents on GitHub: https://github.com/humanlayer/12-factor-agents - SlopCodeBench paper: https://arxiv.org/abs/2603.24755 - SlopCodeBench leaderboard: https://snorkel.ai/leaderboard/slopcode-bench/ - Agentic AI Foundation: https://aaif.io/ TIMESTAMPS 00:17 State of the software factory 00:20 Who is writing 100% of their code with AI, and who still reads it 01:48 Cutting through the loops hype 02:03 The data: review quality down, incidents and bugs up 02:15 Is it a skill issue? 02:29 The software factory since 1968 02:46 Anatomy of a 2022 software factory 04:06 What the agentic software factory actually changes 04:54 The lights-off software factory 05:22 The thesis: unattended models degrade codebase quality 06:13 Why maintainability has no fast oracle 06:29 SlopCodeBench and what it measures 07:00 The best model on the market scored 14.8% 07:23 The lab postmortem that proved the point 08:19 Even Steve came around 09:23 Practice 1: plan before you build 09:59 Do not over-plan, and the expected pain curve 10:58 Practice: align visually with the show me skill 11:31 Anti-slop rules for TypeScript codebases 11:49 Practice: feed user feedback into the factory 12:12 Practice: let models experiment with feature flags 12:34 Loops as forward pressure plus back pressure 12:59 Practice: compound over time from past sessions 13:17 The new software forge and the system of record 14:23 The four layers: compute, dev environment, harness, control plane 15:01 What HumanLayer is building and open sourcing 15:52 Resources and where to find Dex #CodingAgents #SlopCodeBench #AICodeReview