Neura Market
Harness Engineering — OpenAI logo

Harness Engineering — OpenAI

Free

Official OpenAI post: "leveraging Codex in an agent-first world"

FreeFree tier
Type
Open Source
Founded
2015
Company
OpenAI
Visit Website GitHub (40 stars)

About Harness Engineering — OpenAI

This blog post from OpenAI details an internal experiment where a team of engineers built and shipped a software product entirely using Codex agents, with zero lines of manually-written code. Over five months, a small team generated approximately one million lines of code across application logic, tests, CI configuration, documentation, and tooling, achieving an average throughput of 3.5 pull requests per engineer per day. The post explores the shift in engineering roles from writing code to designing environments, specifying intent, and building feedback loops that enable AI agents to work reliably. Key lessons include the importance of underspecified environments, depth-first task decomposition, and that human attention is the scarce resource. The methodology aims to increase engineering velocity by an order of magnitude.

Key Features

Zero manually-written code: every line generated by Codex agents
Agent-first development: humans design environments and specify intent
Depth-first task decomposition: breaking goals into small, agent-actionable blocks
Feedback loops and scaffolding: building capabilities that are legible and enforceable for agents
High throughput: 3.5 PRs per engineer per day with a small team
Continuous iteration: shipping, deploying, breaking, and fixing entirely through agents

Pros & Cons

Pros
  • Estimated 1/10th the time required to write code manually
  • Enables small teams to produce large codebases (millions of lines)
  • Forces focus on system design, scaffolding, and agent capabilities
  • Throughput increased as the team grew, contrary to typical expectations
Cons
  • Early progress was slow due to underspecified environments for agents
  • Requires high-level engineering skill to design agent workflows and feedback loops
  • Dependency on agent capabilities; failures must be addressed by adding missing abstractions, not by trying harder
  • Not suitable for scenarios requiring strict manual oversight or proprietary code constraints

Best For

Building and shipping internal software products with minimal human codingRapid prototyping of new applications using AI-driven code generationScaling engineering output with small teams through agent orchestrationDeveloping automated CI/CD pipelines and observability tooling via agents

FAQ

What is harness engineering?
Harness engineering is a methodology described by OpenAI where software is built entirely by AI agents (Codex), with humans focusing on designing environments, specifying intent, and creating feedback loops rather than writing code.
How did the team build a million lines of code without writing any?
The team used Codex CLI (powered by GPT-5) to generate all code—application logic, tests, CI configuration, documentation, and internal tooling. Humans prompted agents, reviewed pull requests, and improved the environment to enable more complex agent tasks.
What was the biggest challenge in this approach?
The environment was underspecified early on; agents lacked the tools and abstractions to make progress on high-level goals. Human engineers had to identify missing capabilities and make them legible and enforceable for the agents.