Question 1

Everything except the weights

Accepted Answer

System prompts, tool descriptions, tool implementations, middleware, skills, sub-agent configs, and long-term memory—all editable without retraining.

Question 2

Harness vs prompt optimization

Accepted Answer

Prompt tuning edits instructions. Harness engineering edits the full action space—including which tools exist and how trajectories are logged.

Question 3

Start with observability

Accepted Answer

Before auto-evolution: capture traces, classify failures, write change manifests, run regression checks. Every edit explainable, testable, reversible.

Question 4

Transferable harness artifacts

Accepted Answer

AHE shows evolved harnesses transfer across benchmarks and model families—encoding general engineering experience, not overfitting.

Harness Engineering (2025-2026)