Offline coding agent fixed a bug on a laptop in the Himalayas

A coding agent ran entirely offline this weekend on a laptop in the Himalayas. It fixed a bug by itself. Rittika Jindal, a Principal Engineer at Thomson Reuters, wrote the account up for Latent Space. The setting is a hook, but the useful finding is simpler. One of her three tasks worked with no network at all. The model is only 9 billion parameters.

The setup Jindal used

The agent is OpenCode 1.19.4, an open-source coding harness. The runtime is Ollama. The model is Qwen3.5 9B. Jindal says it is 6.6 to 7.6 GB on disk. Ollama's own guidance is 12 to 16 GB of memory. That fits a modern Mac laptop with room to spare.

She set the model to its own context window, roughly 65,000 tokens. She launched Ollama with its cloud mode turned off. She ran OpenCode with the --pure flag, which keeps it inside the project directory. The whole chain works without reaching the network. The model and the dependencies have to be on disk first.

One task worked, two did not

Three exercises were on the plan. An expense calculator, a CSV report generator and a log analyzer. The expense calculator is the one Jindal completed offline. The agent read the project. It found that one line in a Python file was adding instead of accumulating. It fixed the line and ran the tests. All five unit tests passed. Jindal confirmed by running the tests herself after unplugging the network.

The other two tasks she had not completed offline at the time of writing. The post is explicit that one task is one task. A nine-billion-parameter model is a long way from a frontier agent. It will struggle on anything harder than a one-file fix.

What she tells a reader to copy

The lesson is a short pattern for a specific case, which the post describes clearly. The reader is a working developer who knows a project. They need to make a small change on a flight. Or in a conference room. Or at a client site with no network.

Prepare the project and the model while online. Pull the model with Ollama first. Run the exact launch command once with the network up. Download the project's dependencies. Keep the documentation the agent will need on disk. Give the agent the exact test command, so it can tell whether its fix worked. Keep each task focused on one function or one file.

The agent still gets things wrong offline. Jindal recorded both successes and failures. That is the only honest way to know when the pattern is worth trying again. The point is that one of three exercises succeeded with no network at all. The model fits in a laptop's memory.

A team that pays for a frontier model on every ticket now has a free alternative. It only covers a narrow set of tickets. For a team at a client site with locked-down network access, it is a workable path. The kind of task that used to require a trip home now runs in place.

Source

Coding From the Himalayas With an Offline AI Agent, Latent Space, by Rittika Jindal.