Harness Engineering: Leveraging Codex in an Agent-First World
By: OpenAI Type: article URL: https://openai.com/index/harness-engineering/
Summary
OpenAI’s team built and shipped an internal product with 0 lines of manually-written code using Codex agents over 5 months. The product has ~1M lines of code, ~1,500 PRs merged, started with 3 engineers (now 7), averaging 3.5 PRs/engineer/day. Core philosophy: “Humans steer. Agents execute.”
Key topics:
- Redefining the engineer: Primary job became enabling agents to do useful work — designing environments, specifying intent, building feedback loops
- Progressive context disclosure: AGENTS.md as table of contents (~100 lines), not encyclopedia; structured docs/ as system of record
- Repository as the source of truth: Anything not in-repo is invisible to agents (Slack convos, Google Docs, tribal knowledge)
- Agent legibility: Optimize codebase for agent readability first; favor “boring” composable tech
- Enforcing architecture mechanically: Rigid layer model (Types → Config → Repo → Service → Runtime → UI), custom linters, structural tests
- Entropy/garbage collection: Recurring cleanup agents scan for drift, open refactoring PRs; “technical debt is a high-interest loan”
- Increasing autonomy: Single prompt → reproduce bug → fix → video proof → PR → respond to feedback → merge
Key Ideas Extracted
Not yet processed.
Notes
- Context from user: “add to ai-pm / horizontal / software development methodologies”
- Particularly relevant to: agent lifecycle management, context management, software delivery methodology
- “Give Codex a map, not a 1,000-page instruction manual” — aligns with home-brain’s own CLAUDE.md philosophy
- 6+ hour single agent runs while humans sleep
- Agent-to-agent review replacing human review over time
- Custom linter error messages designed to inject remediation instructions into agent context