'/%3e%3cpath%20d='M4%206L17%206L23%2016L17%2022L4%206Z'%20fill='url(%23g)'/%3e%3cpath%20d='M4%206L17%206L20.2%2011.3L15.6%209.9L4%206Z'%20fill='%23fff'%20opacity='0.28'/%3e%3cpath%20d='M44%206H31L19%2024V44H28V26L44%206Z'%20fill='url(%23b)'/%3e%3cpath%20d='M44%206H31L19%2024V44H28V26L44%206Z'%20fill='url(%23g)'/%3e%3cpath%20d='M44%206H31L20.4%2021.8L23.2%2022.9L34.2%208.6H41.6L44%206Z'%20fill='%23fff'%20opacity='0.26'/%3e%3c/svg%3e)
← Theory library
AI Agents & MCP theory
Agent loops, ReAct, tool calling, MCP servers, and multi-agent orchestration.
Practise in the Code Lab →Easy
- 1. What an AI agent actually isAn agent is a loop: model output drives tool calls, tool results go back into the model.
- 2. ReAct: reason, act, observeThe classic prompting pattern that made tool-using agents reliable.
- 3. Tool schemas that models call correctlyNaming, descriptions, and JSON Schema decide whether tool calling works.
- 4. Planner-executor separationOne model call makes a plan; cheaper calls execute the steps.
- 5. Memory: scratchpad, summary, and long-term storeThree distinct memory tiers, each with a different eviction policy.
- 6. Model Context Protocol: why it existsMCP standardises how agents discover and call external capabilities.
Medium
- 7. Writing an MCP serverExpose your own data and actions as standard tools.
- 8. Multi-agent orchestration: supervisor and handoffsSplit work across specialised agents only when a single agent measurably fails.
- 9. Framework landscape: LangGraph, CrewAI, AutoGen, AgnoWhat each framework optimises for, and when plain code is better.
- 10. Sandboxed code executionThe most powerful agent tool, and the most dangerous.
- 11. Human-in-the-loop approval gatesPause before irreversible actions and resume with a durable token.
- 12. Evaluating agents: trajectories, not just answersScore tool choice, step count, and cost alongside final correctness.