Harness engineering: leveraging Codex in an agent-first world
Summary
This blog post details an experiment by OpenAI where a software product was built with “0 lines of manually-written code” using Codex agents. The project involved generating approximately one million lines of code, including application logic, tests, CI configuration, and documentation, in about one-tenth the time it would have taken manually. The core philosophy shifted from manual coding to humans designing environments, specifying intent, and building feedback loops for AI agents to execute.
Extracted Concepts
- Agent-first software development
- Codex as a code generation agent
- Engineering velocity increase
- Redefinition of the engineer’s role (system design, intent specification)
- Repository knowledge as the system of record
- Enforcing architectural invariants
- Legibility for AI agents
- Autonomous development loops
- Entropy management in agent-generated code
Notes Created or Updated
- Agent-First Software Development
- Codex
- Repository Knowledge as System of Record
- Enforcing Architectural Invariants
- Large Language Models