| AI Agents: Tools, Loops and Architectures |
Levels of agency and when not to build an agent; designing tools, schemas and outputs; MCP; the agent loop and managing state; ReAct, CodeAct, plan-and-execute and Reflexion; prompt injection, guardrails, human approval gates and recovering from errors |
Open |
| Lab 1: Build an Agent From Scratch |
No framework: define tools as contracts, parse native tool calls and run the loop with stopping rules, measure how tool descriptions change the model's choices, serve tools from an MCP server, gate the one dangerous tool behind human approval, then attack your own agent with an instruction hidden in an email. |
Notebook Colab |
| Lab 2: Research Agent |
Build a deep-research agent over arXiv in six small functions: the model plans searches, picks papers, reads and summarizes them, and decides when it knows enough; code verifies every citation. Ends with the same agent as a LangGraph graph. |
Notebook Colab |
