Software Engineering as Experimental Practice: OpenAI's Harness Engineering Insights

The OpenAI article on harness engineering ( https://openai.com/zh-Hans-CN/index/harness-engineering/) is a respect-worthy piece because it treats software engineering as an experimental and practical discipline. Verbal debates or concept-stretching analogies (e.g., applying the Tao Te Ching to software engineering) disrespect the field; real progress comes from proposing methods, running experiments, and discovering new insights.

Critique of Hindsight Bias and Narrative Fallacy

Many dismiss the article's techniques as already known, citing harness and OpenClaw as examples. The author invokes the narrative fallacy and hindsight bias from The Black Swan : after an event occurs, we convince ourselves we "should have predicted it," ignoring that true black swans are unpredictable. Everyone falls into the trap of thinking a thing is trivial within their current knowledge, overlooking their own limitations. Predictable technology isn't breakthrough technology — if you can predict it, it's already mature. Therefore, one should not lightly predict outcomes; predictable things are rarely major discoveries.

Why the OpenAI Article Stands Out

The article is humble: it acknowledges AI tools' limitations and the necessity of deep human involvement — organizing knowledge, making agent tools understand context, and keeping humans at the helm while agents execute. A key question posed:

"What capabilities are ultimately needed, and how do we make those capabilities both clearly readable and enforceably executable for agents?"

1. Architecture Executability

Architectural constraints used to live only in documents without enforcement. Later, fitness tools were built to guard architecture, and Google created extensive tooling and engineering platforms to protect Android's architecture. Methods without executability are ineffective. For agents, humans must turn architectural design, decisions, and constraints into code that agents can execute.

The article gives a concrete example: agents can use LogQL to query logs and PromQL to query metrics. With this context, prompts like "ensure service startup completes within 800ms" or

"no span across these four critical user journeys may exceed two seconds"

become feasible.

Performance Execution and Observability Interaction

How performance is enforced and guarded, and how agents interact with observability to achieve performance targets, is presented as a highly practical method.

2. Agent Readability

Traditionally, software and its documentation were optimized for human readability. Now the emphasis shifts to agent readability. The best approach is code that is readable by both humans and agents, which raises the bar for engineers. Converting more system parts into forms that agents can inspect, verify, and directly modify directly increases leverage.

Author's Reflection on AI Code Agents

The author's recent focus is on the difficulty AI Code Agents face in structural creation, construction, and maintenance of complex software — specifically, how to maintain a reliability structure and a performance structure. The OpenAI article offers practical methods for combining software engineering practices to build complex software, with the core insight that human creativity remains crucial.

Conclusion

This is a pioneering exploration of complex software development. Ultimately, companies that earnestly implement software engineering fundamentals — automated testing, observability, knowledge organization, standardized design (e.g., Google) — become more competitive in the AI era. Companies chasing a universal silver bullet will never use AI well nor achieve good results.