Harness engineering
Designing the environment and controls in which an AI agent works
Core idea
Here, harness engineering means AI agent harness engineering: designing and maintaining the instructions, tools, permissions, approvals, records, verification, recovery, and operational observation that help an AI agent work toward an intended goal. It is an emerging umbrella term rather than a standardized discipline, with boundaries that overlap context engineering, agent orchestration, platform engineering, and AI security.
In plain language
It goes beyond telling AI to “be careful.” The working environment identifies the documents to follow, files it may change, operations that need approval, and methods for checking and reversing changes.
Components
- Instructions: Record role, goal, boundaries, and completion criteria in a project-root
AGENTS.mdor another tool-specific instruction file. - Tools: Connect file operations, command execution, browser inspection, and other required capabilities.
- Permissions and approval: Separate destructive, external, paid, or data-changing actions.
- Records and recovery: Use an isolated workspace and version control to checkpoint recorded state. External transmission, payments, API calls, deployments, database changes, and disclosed secrets need operation-specific prevention or compensation and may be irreversible.
- Verification and feedback: Check explicit completion criteria with tests, static analysis, policy checks, logs, state inspection, or human review; retry, stop, or escalate on failure.
- Coordination and observation: Define ownership and handoff formats, and make execution, costs, and failures observable.
Relationship to harness tools and instruction files
A harness tool is a product or component that supplies part of the runtime environment; harness engineering is the activity of designing and operating the full system of instructions, tools, permissions, verification, and recovery. A file such as AGENTS.md is part of a harness only when the agent or host actually discovers and applies it. Writing “do not delete files” does not restrict deletion permission by itself, so sandboxing, approval before execution, secret isolation, checkpoints, and inspection of actual results need to work together.
Relationship to multi-agent work
A harness matters for one agent. With multiple agents, it also defines ownership, shared plans, dependency order, and integration criteria.
Professional terminology
Related terms include agent harness, sandbox, access control, approval gate, audit log, version control, verification loop, and orchestration. Context engineering focuses on information; harness engineering covers the broader environment in which information, tools, permissions, and feedback operate.
Role in this blog
The project folder, AGENTS.md, design documents, OpenCode permissions, Git history, browser checks, and print review together form a small harness for the workshop.