Local coding agents

What you’ll learn

What a coding agent does that a plain assistant does not, how to run one against a local model, and how to do so without letting it loose on your machine unsupervised.

The concept

A coding agent takes a goal, rather than a single instruction, and works towards it: it reads files, plans a series of steps, edits across the codebase, runs commands, and checks results. This is a real step up in capability, and a real step up in what can go wrong, because you are handing a model access to your files and often your terminal.

Two ideas make agents workable:

• Plan then act. Tools like Cline propose a plan and let you approve each step, rather than charging ahead. Keeping a human in that loop is the main safeguard.
• Sandboxing. OpenHands runs tasks inside an isolated Docker container, so what the agent can touch is contained. Running agents in a sandbox, or at least a disposable branch, is strongly advisable.

Agents lean harder on the model than autocomplete does. They need strong tool use and a long context window to hold a codebase in view, so they benefit from a capable local model.

Doing it

1. Use a model suited to agents. Devstral Small was built for agentic coding and pairs naturally with OpenHands; Qwen3-Coder 30B-A3B is a strong, longer-context alternative. Both want a 24GB card for comfortable use.
2. Point the agent at your local model through Ollama or a compatible endpoint, so the whole loop stays on your machine.
3. Start in a contained space. Work on a throwaway branch, or in a sandbox, so nothing the agent does is hard to undo.
4. Supervise, especially early. Watch what it plans before you let it act, particularly anything that touches the terminal.

What can go wrong

• Unsupervised terminal access. An agent that can run commands can do real damage if it misfires. Approve steps, and sandbox the work.
• Confidently wrong changes. Agents can make plausible edits that quietly break things. Review the diff as you would a colleague’s pull request, because that is effectively what it is.
• Losing the thread on large codebases. Even with a long context, an agent can lose track across a big project. Scope tasks narrowly, and give it the relevant files rather than the whole repository.
• Expecting cloud-agent autonomy from a local model. The best cloud agents are more capable. A local agent is a genuine help on feature-sized tasks, not a hands-off replacement for you.

Next steps

If you have not yet, revisit choosing a local coding model to make sure your model is up to agentic work, and see the coding tools catalogue for the agents and their trade-offs.