Field guide / 02
Give an agent a workspace
A coding harness connects a language model to your project, tools, and review process. Here is how to make that connection useful.
Model, agent, harness: what is the difference?
- The model proposes
- It interprets the context and generates text or a request to use a tool. It does not have access to your project unless an application provides that access.
- The agent works through the task
- It uses a model to choose actions, reads the tool results, and decides whether to continue, change approach, or finish.
- The harness runs the process
- It manages context, tools, permissions, and session state. A coding harness can connect the agent to files, a terminal, version control, and a browser.
- You define and review the result
- The editor and diff view help you inspect the work. Tests and a working preview provide evidence that the change meets your request.
Product boundaries vary: an editor can include a harness, and the same agent can appear in several interfaces. Compare the capabilities you need, not just the label on the product.
A useful loop for project work
- Brief. Describe one change, the relevant files, the audience, and how you will know it works.
- Inspect. Have the agent read the existing project and explain the intended approach when the task is complex.
- Build. Keep the work within a reviewable scope. Notice unexpected files, dependencies, or commands.
- Verify. Review the diff, run the relevant checks, and try the changed behavior yourself.
- Keep. Commit the useful change with a clear description. Record any remaining limitation.
Use a disposable practice project first. Review tool permissions before connecting private repositories or external accounts. A successful test checks a specific behavior; it does not establish that every part of an application is correct.
Explore the tools and their documentation
These are examples to investigate, not a ranking. Features, access requirements, and pricing change; use the official documentation for the version you are working with.
- Cursor Agent
A coding agent that combines instructions, a model, and tools for searching a codebase, editing files, and running commands.
- DeepSeek Harness
A plugin-based agent workspace. Explore the official preview and inspect how tools, models, and execution details fit together.
- Google Antigravity
Documentation for its app, command-line, and IDE experiences. Start with the overview to choose a suitable interface.
- Claude Code
A coding agent available through terminal, editor, desktop, and browser interfaces, with file editing and command execution.
- Model Context Protocol
The open protocol used by many AI applications to connect to external data and tools.