Skip to content

Program Synthesis With LLLM

LLLM ships with a Jupyter-based sandbox and a (work-in-progress) Computer Use Agent so you can iteratively synthesize and execute code under tight supervision.

  • Use lllm.sandbox.JupyterSession to spin up a reproducible notebook workspace. The inserted init cell exposes CALL_API, giving your prompts access to the same proxies/tools configured in lllm.toml.
  • Capture tool output via prompt interrupt handlers so the agent can reason about code execution results before deciding whether to continue editing.
  • For browser-driven or GUI-heavy workflows, the lllm.tools.cua module demonstrates how to normalize keyboard/mouse actions and surface screenshots back to the agent loop.

Combine these components with structured prompts (see the neuro-symbolic guide) to build agents that can write, run, and validate code with minimal boilerplate.