“Humans steer. Agents execute.” OpenAI’s Codex team built a production product — a million lines of code across application logic, infrastructure, tooling, and docs — with zero manually-written code in five months. Three engineers drove ~1,500 merged PRs at 3.5 PRs per engineer per day.
The insight isn’t “use AI to code faster.” It’s that engineering becomes a different discipline entirely. Building software still demands discipline, but “the discipline shows up more in the scaffolding rather than the code.” The harness — architecture constraints, documentation, tooling, feedback loops — is the actual product of engineering work.
Documentation becomes the system of record, structured as a map rather than an encyclopedia. A short AGENTS.md (~100 lines) serves as a table of contents pointing to deeper sources in a structured docs/ directory — the same principle behind CLAUDE.md should be a routing table, not a knowledge base, where CLAUDE.md works best as a minimal IF-ELSE directory rather than a monolithic knowledge dump. Knowledge in Google Docs, Slack, or people’s heads “effectively doesn’t exist” to Codex. This is Spec files are external memory that survives context resets taken to its logical extreme — not just specs surviving context resets, but the entire knowledge base being agent-navigable by design. It’s also why Files are the universal interface between humans and agents: the harness runs on files the agent can read.
Architecture becomes constraint rather than art. Rigid layered dependencies, enforced by custom linters, encode “taste” directly into tooling. The resulting code “does not always match human stylistic preferences, and that’s okay” — as long as it’s correct, maintainable, and legible to future agent runs. This is Declarative beats imperative when working with agents at the system level: describe the constraints, let agents satisfy them.
Background Codex tasks function as “garbage collection” — continuously enforcing golden principles rather than accumulating tech debt. Doc-gardening agents scan for stale information. This is Verification is the single highest-leverage practice for agent-assisted coding embedded into the workflow, and it compounds over time through Compound engineering makes each unit of work improve all future work because the harness itself improves with every iteration.
The team estimates they built in 1/10th the time of manual coding. The implication: “give Codex a map, not a 1,000-page instruction manual.” Karpathy’s Autoresearch crystallizes this further: A mediocre agent inside a strong harness outperforms a stronger agent inside a messy one — the harness (metrics, rollback, scoping, observability) determines performance more than the model. The strongest harness with a mediocre agent beats the strongest agent with a messy harness. This extends to how the harness describes itself: Metadata consumed by LLMs needs trigger specifications, not human summaries — when metadata is consumed by an LLM deciding what to invoke, it should specify activation conditions rather than summarize functionality.