Harness engineering: leveraging Codex in an agent-first world | OpenAI
This article strongly resonates and has clear angles for your perspective
Quick Take
This OpenAI case study claims they built a million-line production app with zero human-written code in 5 months, averaging 3.5 PRs per engineer per day. For Brian, this is either the future of his fintech work or marketing fluff - either way, it's worth unpacking the architectural decisions and seeing what's actually transferable to real engineering teams today.
Relevant Domains
Blog Angles
"I Tried OpenAI's 'Agent-First' Development Process (Here's What Actually Works)"
Most of OpenAI's agent-engineering principles are good practices humans should already follow, but their tooling investment is what made the speed possible.
Compare their structured docs approach to his own fintech startup's tribal knowledge problem, or apply their "progressive disclosure" concept to his Chrome extension documentation.
"Why OpenAI's Agent Development Story Doesn't Scale (Yet)"
The OpenAI case study works because they had unlimited compute budget and a team of AI experts optimizing for agents - normal engineering teams should focus on the organizational lessons, not the full automation.
Analyze which parts he could realistically implement at his fintech startup vs. which require OpenAI-level resources.
"The Real Cost of Agent-Generated Code"
OpenAI's approach trades code review time for infrastructure complexity and ongoing "garbage collection" - quantify whether this math works for typical engineering teams.
Compare their 20% Friday cleanup time to his current code review overhead, or relate to technical debt management in his credit-card platform.
"Building Side Projects in the Agent Era"
The agent-development techniques that matter most for solopreneurs are the architectural constraints and documentation practices, not the full automation.
Apply their layered domain architecture to his print-on-demand automation or use their progressive disclosure docs approach for his web agency tools.
Key Quotes
We estimate that we built this in about 1/10th the time
1,500 pull requests... 3.5 PRs per engineer per day
give Codex a map, not a 1,000-page instruction manual
anything it can't access in-context while running effectively doesn't exist
we started encoding what we call 'golden principles' directly into the repository