#harnessed.md
"The primary job is no longer to write code, but to design environments, specify intent, and build feedback loops that allow agents to do reliable work."
— OpenAI, Harness Engineering
##What’s harness engineering?
Engineering has fundamentally changed.
Traditionally engineers spent 80% of their time on features, 20% on the system. Harnessed companies flip that on its head: 80% goes to the harness — the machine that builds the machine — 20% guiding where it goes.
The harness has three parts:
- Guides — steer the agent before it acts: AGENTS.md, design docs, architecture maps, rules, learnings
- Verification — checks the work before it ships: types, linters, tests, agentic review
- Observation — monitors what shipped: errors, usage, agentic investigation
Intent ◄········································╮
│ improvements + fixes ·
▼ ·
Guides ◄································┬···╮ ·
│ evolve · · ·
│ constrain + direct · · ·
▼ · · ·
Agent builds ◄───────┐ · · ·
│ │ · · ·
▼ │ · · ·
Verification ········╁··················╯ · ·
│ │ · ·
▼ no: fix │ · ·
Pass? ───────────────┘ · ·
│ · ·
│ yes · ·
▼ · ·
Ship · ·
│ · ·
▼ · ·
Observation ································┴···╯
The system improves itself. Signals from verification and observation loop back into the guides — you don’t just fix the code, you refine the harness to stop it breaking that way again.
The audit is a rubric your agent runs against your repo — a one-line prompt, a 0–5 score per item, and a recommended next step grounded in the tooling you already use.
##Key reading
- My AI Adoption Journey — Mitchell Hashimoto The origin of harness engineering
- Harness Engineering — Ryan Lopopolo / OpenAI 1M lines of code, zero written by humans
- Harness Engineering for Coding Agent Users — Birgitta Böckeler / ThoughtWorks The guides and sensors framework
- Compound Engineering — Kieran Klaassen, Dan Shipper / Every How Every codes with agents
- Context Engineering for Coding Agents — Birgitta Böckeler / ThoughtWorks Configuring what the agent sees
- Effective Harnesses for Long-Running Agents — Anthropic Making agents remember across sessions
- Effective Context Engineering for AI Agents — Anthropic The smallest set of high-signal tokens
##Companies
Software engineering has radically changed. These companies have seen it early and gone all-in on harness engineering. They're hiring.