Give an LLM a regular UNIX environment and a workspace, and let it run shell commands until it completes some task. Runs in a Docker container, edits files in workspace/
and reads tasks and prompts from system/
.
Example tasks:
- system/task_build_run.md - build and run a tiny program in Go
- system/task_python_fibonacci.md - Python script that outputs the first n Fibonacci numbers
- system/task_wikipedia_title.md - download the Wikipedia homepage and parse the title
Task is passed as an argument to the simple-agent
binary in compose.yml
.
This is for quickly experimenting with what LLMs are able to achieve with only straightforward shell access rather than well-defined tools specified ahead of time like ChatGPT functions. In the future, a more general AI should be able to use just a regular command line to achieve a lot of tasks.
- Ollama
- Docker
- able to run the qwen2.5-coder:32b-instruct-q5_K_M model - it may work with smaller Qwen2.5-Coder instruct models
ollama pull qwen2.5-coder:32b-instruct-q5_K_M # if you use a different model, also update the compose.yml
docker compose build
docker compose up # to use a different task, update the compose.yml