Documentation
¶
Overview ¶
Package main implements a CodeTool that lets the LLM write and execute code inside an agent run. The tool contract (name, parameters, SandboxRuntime interface) is sandbox-agnostic — execution is handled by whatever SandboxRuntime is wired in main.go.
local_runner.go — default runner using os/exec (needs Python or Node installed locally) docker_runner.go — isolated runner using Docker (SANDBOX_ENV=docker)
Package main — SandboxRuntime backed by Docker.
This file is the Docker-specific execution layer for the agent_with_code_execution example. It satisfies the SandboxRuntime interface defined in code_tool.go.
To use a different sandbox (AWS Lambda, Modal, E2B, Firecracker, etc.), add a sibling file implementing SandboxRuntime and wire it in main.go instead of newDockerRunner.
Requires: Docker daemon running and internet access to pull language images on first run.
agent_with_code_execution demonstrates sandboxed code execution inside an agent.
The goal: the LLM writes code to answer a user question; the code runs in an isolated sandbox and returns the output. The LLM reads the output and replies. The sandbox is pluggable — swap the runtime without changing the tool contract.
code_tool.go — CodeTool and SandboxRuntime interface (sandbox-agnostic) local_runner.go — default runtime using os/exec (needs Python or Node on host) docker_runner.go — isolated runtime using Docker; activate with SANDBOX_ENV=docker
Run (local sandbox — needs Python or Node installed):
go run ./agent_with_code_execution "Write a Python script that prints the first 10 Fibonacci numbers"
Run (Docker sandbox — needs Docker daemon):
SANDBOX_ENV=docker go run ./agent_with_code_execution "Write a Python script that prints the first 10 Fibonacci numbers"
Source Files