Model, Harness, Tools and Workspace

Definition

Model, Harness, Tools and Workspace

A coding agent’s capability is the product of four layers: the model that reasons and generates text, the harness that manages state and execution, the tools that expose concrete actions, and the workspace that supplies files, environment, credentials and repository history. Changing any layer can change the agent’s behaviour, risk profile and evidence quality even when the user prompt stays the same.

Why It Matters

The model supplies language understanding, code generation and reasoning. Model capability matters, but does not determine the agent’s complete behaviour.

The harness provides tool execution, file access, state, context assembly, approvals, sandboxing, progress, retries and stopping conditions. Different harnesses can produce substantially different outcomes with similar models.

Tools are controlled capabilities such as repository search, file editing, shell execution, browser access, issue tracking or deployment. The workspace model is the files, repository state, environment and connected resources visible to the agent.

Failure mode

The recurring risk is treating this concept as a label rather than an engineering decision with trade-offs, evidence needs and operational consequences.

Evaluation check

Ask what observable evidence would show that this concept has been applied well in a real system, repository or learning exercise.

Related Concepts

• Tool calling and MCP
• Configuration, permissions and sandboxing

Further study: OpenAI, Codex customization, and Anthropic, Model Context Protocol.