← back
Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI
Original: Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI
Takeaway
Treat agents as infinitely-parallel implementers and invest your time in writing guardrails and persona-oriented docs that make 'a good job' legible to them.
Summary
- Lopopolo's team at OpenAI bans editors and works exclusively through agents; each engineer commands 5-5000 parallel 'teammates' bounded only by tokens and GPUs.
- Claims GPT-5.2 reached isomorphic-to-human software engineering ability; code is now 'free' and the scarce resources are human time, attention, and model context window.
- Large-scale migrations and P3 tickets that used to languish can be parallelized across 15 agents; localization/i18n becomes free.
- Write down non-functional requirements (QA plans, style, ADRs) as durable agents.md guardrails so every agent trajectory inherits the team's best practices.
- GPT-5.4 Codex autocompaction is good enough that he never types /new; staff-engineer skill set shifts to systems thinking and delegation.
agentscoding-agentsharness
Original description
https://openai.com/index/harness-engineering/ Speaker info: - https://x.com/_lopopolo - https://www.linkedin.com/in/ryanlopopolo/ - https://github.com/lopopolo With a special post keynote Q&A with Vibhu Sapra (https://x.com/vibhuuuus), cohost for https://latent.space/p/harness-eng