🔧 Harness Engineering - the new discipline you should know about
Models are commoditized. Claude, GPT, Gemini - they all perform within a narrow band of each other on benchmarks. The competitive advantage has shifted to what you build around the model.
That's harness engineering.
A harness is the complete system that wraps an AI agent. It governs:
- 🛠️ What the agent can access - tool permissions, file system scope, API access
- 📋 What it should do - context engineering, prompt structure, instructions
- ✅ Verification - automated checks, test runners, lint, build feedback
- 🔁 Self-correction - feedback loops that let the agent fix its own mistakes
- 🙋 Human-in-the-loop gates - approval checkpoints for high-stakes actions
The discipline only entered mainstream use in the last couple months, but the implications are significant.
OpenAI's Codex team shipped a production codebase with over 1 million lines of code. Zero lines written by humans. The engineers didn't write code - they designed the harness that let the agent write it reliably.
The key insight: the model is not the product. The system around it is.
This is also closely tied to context engineering - what information the agent has access to at each step determines whether it behaves coherently over long tasks.
Worth understanding deeply if you're building anything with AI in 2026. 👇
What sources are you following for this? Drop them in the comments.
More notes on engineering, AI and the systems around the work.
Back to all writing