Harness engineering becomes agent battleground

The useful shift here is that “agent engineering” is becoming less about clever prompts and more about boring systems work: constraints, observability, verification, permissions, and recovery.

• –OpenAI’s version is the most aggressive: make the repo itself legible to Codex, encode taste as tooling, and let agents iterate through PRs, tests, and reviews.
• –Anthropic’s answer is more platform-shaped: decouple the model from the tools and runtime so managed agents can run longer jobs while the harness evolves underneath.
• –ThoughtWorks brings the enterprise lens: treat harnesses as feedforward guides and feedback sensors, mixing deterministic checks with AI review.
• –For developers, this makes harness design a new leverage point: better tests, linters, permissions, docs, and state handling can outperform simply swapping models.