Build · Advanced
Agents, tools and MCP
Tool-using models, agent loops, the Model Context Protocol and how to keep autonomy bounded.
Why it matters
Agents turn a model into something that acts. The hard parts are permissions, approval steps and recovering when a step goes wrong.
Key ideas
- Function and tool calling
- Reason-and-act loops
- Model Context Protocol servers and clients
- Human approval steps
- Failure recovery
You will be able to
- Write an MCP server
- Design least-privilege tools
- Add approval gates to an agent
Primary reading
- ReAct: Synergizing Reasoning and Acting in Language Models (opens in a new tab)Yao et al. · 2022
- Toolformer: Language Models Can Teach Themselves to Use Tools (opens in a new tab)Schick et al. · 2023
- Model Context Protocol specification (opens in a new tab)MCP project · 2024