Lab: The bounded Cases agent over MCP

Module 9 · Lab · About 75 minutes · Needs: lessons 1 to 6 of this module, and the Module 8 lab (or nothing: Step 1 rebuilds Modules 1 to 8)

What you are building

Maya Ortiz's monitor from Lessons 1 and 6 now runs on the whole platform. The six lessons' files go into the repository unchanged and are joined into one Cases agent: Lesson 1's bounded LangGraph loop calls its tools through four MCP servers behind Lesson 3's gateway; each run is an RQ job that pauses for approval with a checkpoint and resumes on any worker; every call is a span and a line in the hash-chained audit trail; kill switches are read before every call; and trajectory evaluation scores each run's path.

The eight Cases tools are per-intent tools, as Lesson 2 names them. One optional cell calls your provider; every number quoted comes from scripted models, in-memory MCP and fakeredis. Open a new Colab notebook, module-09-lab.ipynb, and run each cell in order.