Interview Question: What Is Harness Engineering and How to Answer It

Standard Answer for the Interview

Harness Engineering, also known as "驾驭工程," is a systematic approach to designing the work environment for AI agents. It includes context management, tool invocation, sandboxing, permission control, testing, logging, review mechanisms, and feedback loops, aiming to make agents more reliable, controllable, and capable of continuously completing tasks.

Problems Harness Engineering Solves

Traditional coding with large models often involved a one‑off interaction: describe a requirement, generate code, copy it, and manually judge its usefulness. Modern coding agents (e.g., Claude Code, Codex) read repositories, modify files, run commands, handle errors, and submit PRs, working continuously. However, they still encounter issues such as undocumented architectural rules, unclear project startup procedures, scattered test entry points, missing logs, UI verification requiring a browser, and non‑mechanized code‑review standards. These gaps cause even strong models to make low‑level mistakes.

The core idea of Harness Engineering is to provide agents with proper environmental conditions:

When context is missing, structure it.

When tools are missing, expose them.

When feedback is missing, integrate tests, logs, reviews, and execution results.

When boundaries are missing, add permissions, architectural constraints, and secure sandboxes.

Components of Harness Engineering

Harness can be divided into two parts: guidance and feedback.

Guidance (pre‑task): Documents and artifacts that tell the agent how to act, such as AGENTS.md, architecture documents, coding standards, project startup scripts, API docs, domain knowledge, example tasks, skills, and task decomposition methods.

Feedback (post‑task): Signals that tell the agent whether it succeeded, including unit tests, type checks, linters, static analysis, architecture tests, browser automation, runtime logs, metrics, traces, code‑review agents, LLM judges, and human reviews.

A good Harness enables the agent to operate in a loop of "understand task → make changes → observe feedback → fix issues → re‑validate."

Difference Between Harness Engineering and Prompt Engineering

Prompt engineering focuses on input design to get a better immediate model response, while Harness engineering focuses on system design to ensure long‑term reliable operation of the model within an engineering environment.

Example: Adding "Please follow project architecture, do not cross‑layer calls" in a prompt is prompt engineering. Embedding the same rule in a custom linter that automatically checks every code change is Harness engineering.

Similarly, asking the agent to "run tests after modification" in a prompt is still prompt engineering; intercepting the agent before it finishes, forcing it to execute tests, capture full output, and loop back on failures constitutes Harness engineering.

Prompt engineering can be a component of Harness engineering, but they are not opposites.

Practical Harness Implementation Example

To illustrate a real Harness, consider a coding agent used in production:

Use AGENTS.md as the entry point to describe project structure, common commands, and key constraints.

Place detailed architecture documents in a docs directory for the agent to read on demand.

Apply linters and architectural tests to enforce layer boundaries.

Provide local startup scripts and test commands.

Expose logs, metrics, and traces so the agent can observe execution results.

Integrate Playwright for front‑end tasks, allowing the agent to open pages and verify UI.

Add a final checklist requiring the agent to confirm requirements, test outcomes, and modification scope before finishing.

Introduce a dedicated review agent or human review for complex tasks.

The ultimate goal is not to replace engineers but to surface implicit engineering knowledge—architecture rules, testing standards, security constraints—to the AI, freeing engineers to focus on higher‑value work. Harness engineering is thus considered a fundamental skill for the AI‑agent era.