Harness Engineering
/ˈhɑːrnəs ˌendʒɪˈnɪərɪŋ/
Definition
Designing and optimizing the "harness" around an LLM or AI agent — its tools, prompts, action space, observation formatting, verification loops, and permissions — to get higher task-completion rates from the same model.
Examples
- • Before swapping models, we did harness engineering and cleaned up the tool definitions first.
- • Good harness engineering absorbs an agent's failures with observe-and-recover loops.
Origin
Emerged as agentic AI spread and the system *around* a model — not the model alone — increasingly determined performance.
A "harness" originally means the gear that safely holds and steers something. In AI it refers to the whole system around a model that makes it actually useful — which tools it gets, how observations (results) are formatted, what gets verified and when to retry, and what permissions and guardrails apply.
Core activities:
- Action-space design: define tools/APIs so the agent doesn't get confused
- Observation formatting: shape results and errors for the next decision
- Verification loops: pair with iteration so the agent self-checks and recovers
- Evaluation (evals): measure whether a harness change actually raised completion rate
The same model can perform very differently with a better harness, so improve the harness before scaling the model often wins.