What it is and who makes it
A harness is the system around a model that gives it context, tools, permissions, an execution loop and verification.
Maker or project: Concept guide — no single vendor. Classification: Concept guide; this page treats the surrounding context, tools, permissions and verification as the relevant harness boundary.
The model produces a decision or tool call from the current context.
The harness supplies tools, execution environment, context files, permissions and session state.
The harness executes approved actions and feeds results back into the model loop.
Verification, limits and human review decide whether the work is accepted.
Who it may fit
Pick an agent by its control surface and trust boundary, not only by its model label.
Useful scenarios
- Developers evaluating AI coding tools
- Teams designing safe agent workflows
- People confused by agent, harness, gateway and orchestrator terminology
Trade-offs and checks
- Harness is not a universally standardized product category; ask what the tool actually controls rather than trusting the label.
- A stronger model does not compensate for missing tests, unsafe permissions or bad context.
- An orchestrator can plan work while a harness performs actions; combining the names does not remove either responsibility.
Interactive vs headless and CI
These modes change the approval, identity and observability boundary.
There is no universal harness command. Interactive sessions, background workers and CI should all use an explicit workspace, tool allow-list, approval policy, timeout and verification contract.
Request-to-verification workflow
Use this mental model when debugging a surprising edit, tool call or cost.
Intent + context
Your prompt, repository instructions, selected files, history and environment become model context.
Tool + permission
The harness proposes a read, edit, shell, web, MCP or other action; policy decides whether it can run.
Result + verification
Command output, compiler errors, tests and your review return to the loop before acceptance.
Setup outline
Start with the smallest safe experiment, then follow the current official documentation for the real installation.
Model + context files and repository map + tools: read, edit, shell, test + permissions and approval gates + session and memory + verification loop = coding-agent harness Gateway/router: chooses where model traffic goes. MCP: standardizes one kind of tool/context connection. Orchestrator: coordinates one or more agents.
Instructions, skills, MCP and hooks
These surfaces can shape context or add actions; they are part of the trust boundary.
There is no universal filename or plugin contract. Inventory repository instruction files, skills, plugins, MCP servers and hooks as executable configuration; document precedence, trust and review ownership before enabling them.
Official sources
Use these sources for current installation, capabilities, plans, releases and security notices.
Reviewed 2026-09-26. This date is a change-tracking note, not a promise that the product remains unchanged.
Related learning
Related agent guides
Want the foundations behind the workflow?
Study AI fundamentals, prompt design, AI-assisted development, agents and MCP without a forced linear course.