LLM Wiki

Appearance

Built 26/04/17 09:08commit f8ff6f9

OpenAI Harness Engineering In An Agent-First World ​

中文 | English

Summary ​

This source argues that once agents write nearly all code, the engineering bottleneck shifts from manual implementation to harness design, repository legibility, and feedback loops that let agents operate reliably with minimal human attention.

Source ​

Key Contributions ​

  • Reframes software engineering as environment design: humans steer, specify intent, and encode feedback loops while agents execute.
  • Treats repository-local documentation and plans as the system of record, with AGENTS.md serving as a map rather than an encyclopedia.
  • Emphasizes legibility to agents across UI behavior, logs, metrics, traces, architecture boundaries, and review workflows.
  • Argues that architecture constraints and custom lints are multipliers because they convert human taste and invariants into mechanically enforced guidance.
  • Connects high agent throughput to a different merge philosophy: cheap correction loops, short-lived PRs, and more review shifted agent-to-agent.

Strongest Claims ​

  • A fully agent-generated codebase can remain productive only if the repository itself exposes enough structure, tooling, and documentation for agents to reason directly from local context.
  • Human effort compounds best when invested in scaffolding and control systems rather than direct code writing.
  • Monolithic instruction files fail; progressive disclosure through a short map plus indexed deeper docs scales better.
  • Agent autonomy introduces entropy, so recurring cleanup passes and enforceable golden principles are required to prevent slop from compounding.

Practical Implications For This Vault ​

  • Keep AGENTS.md short enough to orient the model and push domain detail into linked canonical pages.
  • Treat topic pages and source pages as agent-facing infrastructure, not just human notes.
  • Prefer repository-local durable artifacts over relying on chat history or external documents.
  • Lint and ingest work should watch for places where structural guidance belongs in canonical pages rather than in transient conversation.