Codegen agents are becoming standard for both local and remote workloads. As their capabilities grow, so does the blast radius when their experiments go wrong: deleted files, leaked secrets, corrupted state, and in adversarial cases, host compromise.
Sandboxes are the emerging answer. Think of them like a laptop for agents: close the lid and the environment suspends, open it and it resumes instantly with full state preserved. Hardware-level isolation with sub-second startup, giving agents somewhere to run arbitrary code without touching the host.
With a sandbox in place, the next question is what to put in front of it. A single agent with a big toolbox and one sandbox is the obvious starting point…and it works, until it doesn’t. Context windows balloon, the model second-guesses its own tool choices, and one prompt ends up trying to do reviewer, refactorer, and test-writer all at once.
Specialized agents are a better unit of composition: each one narrow in scope, easier to evaluate, easier to swap out, and easier to reason about when something goes wrong. But the hard part isn’t building one agent; it’s getting two of them to share enough context to collaborate without stepping on each other.
This full-day, hands-on workshop walks attendees through the complete journey: from a single-prompt prototype to a sandboxed, multi-agent system with shared context. We will build it using the OpenAI Agents SDK (Apache 2.0, LLM-agnostic) wired to a cloud sandbox via MCP. The pattern is vendor-agnostic and works with any LLM and sandbox provider supported by the SDK. Teams with data sovereignty requirements can swap in self-hosted open-weight models such as Llama 3 or Mistral served via vLLM or Ollama.