Harness Engineering: Leveraging Codex in an Agent-First World

engineering blog · source date 2026-02-11 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Codebases are written to be legible to humans, which is not the same as being legible to a coding agent.
When agents write most of the code, the bottleneck moves from writing to specifying, validating, and observing.
Agent failures in a large repo are hard to diagnose without per-run observability.

A case study of an OpenAI team shipping roughly 1M lines and about 1,500 PRs over five months with zero human-written source.
The mechanism is making the repo legible to Codex: an AGENTS.md map, structural tests, linters that teach the remediation rather than just flagging, DevTools/observability integration, ephemeral per-worktree observability (LogQL/PromQL), and background cleanup agents.
Reframes the engineering job as designing environments, specs, and feedback loops rather than writing implementations.

Alongside Anthropic's harness-design post, this is one of the two most-cited practitioner anchors for harness engineering, and it is repeatedly referenced by later papers (The Last Harness, Self-Harness) and blogs.
The teaching-linter idea generalizes: an error message an agent can act on is worth more than one a human can interpret.
Caveat: a vendor account. The ~1M-LOC and zero-human-source figures are OpenAI-reported and unaudited, and this describes an aspirational internal workflow rather than a reproducible benchmark.

Comments

No comments yet.