Code Sandbox Tech Behind Manus and Claude Agent Skills (www.dataleadsfuture.com)

This tutorial shows how to build a stateful, self-hosted Jupyter code sandbox that reproduces the core tech behind commercial products like Manus and Claude Agent Skills—letting LLM agents generate and run Python interactively instead of relying on stateless command-line executors or fixed function-calls. The author provides a reproducible stack: a custom Docker image (python:3.13-slim-bookworm) running jupyter_kernel_gateway (exposed on port 8888 with an auth token), a DockerJupyterServer that starts the container, a DockerJupyterCodeExecutor that uses the Jupyter Kernel API, and an Autogen CodeExecutorAgent to submit code. You install autogen-agentchat and autogen-ext[docker-jupyter-executor], build the jupyter-server image, and then run examples showing persistent kernel state (define x in one message, read x in the next), which stateless sandboxes cannot do. Significance: a stateful Jupyter backend dramatically improves agents’ ability to inspect unknown data, iterate like human analysts, and solve complex math/data tasks more cheaply and flexibly than paid code-execution services. The guide also covers production realities—limits of “Docker out of Docker,” why agents should connect to pre-deployed Jupyter servers (e.g., on vast.ai) for large RAM/GPU workloads, Docker Compose deployment, idle-resource reclamation tweaks, and integration paths for other frameworks like LangChain—enabling secure, scalable, enterprise-grade agent sandboxes without vendor lock-in.