Track 4 · Agents · Purple
AI agents & orchestration
An agent is not a smarter chatbot—it is an LLM inside a bounded loop that can call tools, retain working memory, and plan multi-step tasks until a termination condition fires. This track teaches how to wire native tool-calling APIs, validate arguments before side effects, implement ReAct-style reasoning, compose multi-agent graphs, plug MCP servers, and wrap RAG as one capability among many. Assumes Tracks 1–3: you can call an LLM API, run a RAG pipeline, and ship versioned prompts with eval gates.
Guides in this track
Six deep-dive chapters. All guides are live—read in order.
Reading order: Agents explained → Tool calling → Agent memory → Multi-agent orchestration → MCP explained → Agentic RAG
-
01
AI agents explained
What agents are, why they work now, agent vs chatbot, observe→think→act loop, ReAct pattern, failure modes, production checklist.
-
02
Tool calling
Function schemas, allowlists, argument validation, parallel tool calls, timeouts, and safe execution routers in Python and Java.
-
03
Agent memory
Working memory in the loop, session summaries, long-term stores, checkpointing, and what to persist vs trim each round.
-
04
Multi-agent orchestration
Supervisor patterns, handoffs, LangGraph state machines, human-in-the-loop interrupts, and when one agent beats many.
-
05
MCP explained
Model Context Protocol servers, tool discovery, auth boundaries, and plugging external systems without bespoke SDK sprawl.
-
06
Agentic RAG
When the agent decides to retrieve, query reformulation, multi-hop evidence, and combining read tools with write guardrails.