Agent Engineering Study Plan
A 12-week curriculum covering the full agent stack — the ReAct loop, planning, memory, tool use, multi-agent frameworks, computer-use and voice agents, prompt-injection defense, and eval-driven development.
Recommended Study Path
Prerequisites
LLM Study Plan
- Transformer architecture
- Prompting & in-context learning
- Function calling basics
Familiarity with the LLM study plan is recommended →
Loop & Planning
Weeks 1-4
- Ch 1-2: ReAct & Planning
- Ch 3-4: Reflexion, ToT, Tools
The agent loop, decomposition strategies, and tool use
Memory & Frameworks
Weeks 5-8
- Ch 5-6: MemGPT, Mem0, Voyager
- Ch 7-8: LangGraph, AutoGen, CrewAI
Long-term memory and the major orchestration frameworks
Production
Weeks 9-12
- Ch 9-10: Computer Use, Voice, Coordination
- Ch 11-12: Injection Defense, Eval-Driven Dev
Specialized agents, safety, observability, and benchmarks
All Chapters
The Agent Loop & ReAct Foundations
What an agent is, the perceive-think-act loop, the ReAct pattern, what early agents got wrong, and the modern agent design space.
Planning Patterns: Plan-and-Execute, ReWOO, HTN
Decoupling planning from execution — Plan-and-Execute, ReWOO's DAG planner, hierarchical task networks, and when to plan up-front vs react.
Self-Improvement: Reflexion, Self-Refine, Tree-of-Thought, LATS
Verbal reinforcement learning, self-critique loops, tree-structured reasoning, and Language Agent Tree Search.
Tool Use & Function Calling
Function calling fundamentals, JSON-mode and tool schemas, parallel tool calls, the Model Context Protocol, and tool-result handling.
Memory I: Virtual Context & MemGPT
The context-window problem, MemGPT's OS-inspired paging, memory blocks, sleep-time compute, and memory promotion.
Memory II: Hybrid Stores, Mem0 & Skill Libraries
Vector + graph + lexical hybrid retrieval, the Mem0 architecture, Voyager-style skill libraries, and how to evaluate memory quality.
Workflow Patterns & LangGraph Stateful Graphs
Anthropic's workflow taxonomy — chaining, routing, parallelization, orchestrator-worker — and LangGraph's stateful-graph runtime with checkpointing.
Multi-Agent Frameworks: AutoGen, CrewAI, OpenAI Agents SDK, Claude Agent SDK
AutoGen's actor model, CrewAI roles, OpenAI Agents SDK, Claude Agent SDK with subagents, the Agno and Mastra runtimes, and a framework selection matrix.
Specialized Agents: Computer Use & Voice
Computer-use agents that drive a screen, action grammars and grounding, Pipecat voice pipelines, LiveKit streaming, and latency budgets for realtime.
Multi-Agent Coordination & Failure Modes
Multi-agent debate and consensus, chairperson moderation, the failure taxonomy of agentic systems, cascading tool failures, and orchestration patterns for coordination.
Safety: Prompt Injection & Defense
The prompt-injection taxonomy, direct vs indirect attacks, data tagging and trust boundaries, output validation firewalls, and red-teaming agents.
Production: Observability, Benchmarks, Runtimes & Eval-Driven Development
SWE-bench, GAIA, WebArena, and OSWorld benchmarks, OpenTelemetry GenAI conventions, observability platforms, production runtimes, and the eval-driven development loop.
Practice Problem Sets
Sharpen your skills with coding challenges and system design problems.
Curriculum designed to take you from the simplest ReAct loop to deploying coordinated, observable, hardened multi-agent systems in production.