Deploying code agents without all the agonizing pain

Agents that write and run code are powerful, as Cognition Labs showed with their recent release of Devin, the "AI SWE". But they are complex to program, hard to deploy, and even harder to secure -- what happens if your agent runs DROP prodtables or sudo rm -rf /? In this joint webinar between LangChain and Modal Labs, we cover the productionization of a coding agent. Lance Martin (@rlancemartin) walk through his coding agent implementation, which performs import and code execution checks along self-reflection in LangGraph. Modal AI Engineer Charles Frye (@charles_irl) will then show how to secure that prototype agent using Modal Sandboxes and deploy it as a FastAPI web app with only a dozen more lines of code. Slides: https://docs.google.com/presentation/d/1368-i3k73eM-h1vsd0LwchxQOC8JUQt7RRy9b44EBho/edit?usp=sharing Code: https://github.com/modal-labs/modal-examples/tree/main/06_gpu_and_ml/langchains/codelangchain First video discussing the design of the self-corrective coding agent in detail: https://www.youtube.com/watch?v=MvNdgmM7uyc Try Modal! Includes $30/month of free compute: https://modal.com Timestamps - 00:00 Summary 00:48 From paper to notebook 04:09 Evaluating the agent 08:20 From notebook to production - LangServe and Modal.asgi_app 13:20 Notebooks and apps 16:45 Iterating in production - OpenAPI docs 18:07 Securing code agents with Modal Sandboxes 23:47 Development servers with modal serve 28:42 Serving a UI with LangServe Playground 37:33 Deeper dive on using Modal Sandboxes 42:20 Observability and monitoring with LangSmith 45:08 Recap (edited)