A coding agent consists of a model plus the surrounding harness that manages prompts, tools, policies, and recovery paths. Harness engineering treats that scaffolding as a concrete artifact, tightening it each time the agent errs. The practice urges engineers to codify failures as rules—adding prompts, hooks, or sandbox constraints so the same mistake does not recur.

Viv Trivedy coined the term and outlined the components: system prompts, skill files, sub‑agent prompts, tool descriptions, bundled infrastructure such as filesystems and sandboxes, orchestration logic, middleware hooks, and observability layers. Simon Willison reduces the loop to “run tools in a loop to achieve a goal,” emphasizing that the design of tools and the control loop carries most of the performance weight.

Anthropic’s engineering team published a detailed harness design for long‑running work, and Human Layer frames most agent failures as configuration issues rather than model deficiencies. Data from Terminal Bench 2.0 shows Claude Opus 4.6 scoring far higher in a custom harness than in its default environment, moving from the top‑30 to the top‑5.

The discipline recommends building harnesses backward from desired behavior, treating each component as a targeted solution to a documented failure. This shifts focus from awaiting better models to extracting capability through tighter prompts, stronger back‑pressure, and safer execution environments.