06 · Agents & Orchestration
Multi-step autonomous systems. Concepts first, frameworks second — understand the patterns before picking the tools.
| Step | Topic | One-liner | Status |
|---|---|---|---|
| 1 | AI agent fundamentals | What makes something an agent: perception, reasoning, action, memory | 🔴 |
| 2 | Tool / function calling | How LLMs invoke external functions and interpret their results | 🔴 |
| 3 | ReAct pattern | Reason → Act → Observe loop — the foundational agent architecture | 🔴 |
| 4 | Chain of Thought | Prompting the model to reason step-by-step before answering | 🔴 |
| 5 | Plan & execute | Separating planning from execution for complex multi-step tasks | 🔴 |
| 6 | Human-in-the-loop | Patterns for routing agent decisions to humans for approval | 🔴 |
| 7 | MCP — Model Context Protocol | Standardized protocol for agents to discover and call tools | 🔴 |
| 8 | Agent SDK | SDK-level primitives for building agents (Anthropic, OpenAI, Google) | 🔴 |
| 9 | LangGraph | Graph-based agent orchestration with explicit state and control flow | 🔴 |
| 10 | LangChain & LlamaIndex | The two dominant agent/RAG frameworks and when to use each | 🔴 |
| 11 | Multi-agent systems | Networks of specialized agents collaborating on shared tasks | 🔴 |
| 12 | Handoff — agent to agent | Patterns for transferring control between agents cleanly | 🔴 |
| 13 | Tool registry | Centralised catalog of available tools with schemas and permissions | 🔴 |
| 14 | Idempotent tool calls | Making tool calls safe to retry without side effects | 🔴 |
| 15 | A2A protocol | Google's agent-to-agent communication protocol (2025) | 🔴 |
← Previous section: 05 · Retrieval & Memory | Next section → 07 · Safety, Alignment & Governance