A week-by-week plan · AI Engineering
Production Agents
For engineers putting AI agents in front of real users, with real tools and real data. You can build an agent that calls tools in a loop, stays inside a budget, leaves a trace you can read, and asks a human before a risky step.
- 5 steps
- 25 lessons
- 2 free
- 1 interview guide
- 1 mock interview
- 27 hours of reading
Who it is for
Engineers putting AI agents in front of real users, with real tools and real data.
What you can do at the end
You can build an agent that calls tools in a loop, stays inside a budget, leaves a trace you can read, and asks a human before a risky step.
First milestone
A tool loop, a governor (limits on steps, time and money), tracing, and human approval.
Milestone: Design a controlled agent
2 of the 25 lessons are free to read now, with no account. The other 23 are in the AI Engineering course: $49 once, lifetime access, pay with PayPal.
Get AI Engineering · $49Sign up free and this page ticks off the lessons you finish and shows you where to carry on.
The plan, week by week
Times are each lesson’s own reading time, added up. A mock interview is set to 45 minutes. A lock means the lesson is in the course.
- Week 1
The tool loop
233 minSee why agents fail as tasks get longer, and what a tool call really is.
- The Agent Loop: Why AI Agents Fail as Tasks Get LongerAI Engineering lesson42 minfree
- Tool Calling Is a Contract, and the Model Will Break ItAI Engineering lesson39 minin the course
- Agent Planning: ReAct vs Plan-and-ExecuteAI Engineering lesson39 minin the course
- Agent Memory and Context: What the Model Sees, and What It ForgetsAI Engineering lesson45 minin the course
- Too Many Tools, or Tools That Look Alike? Measuring How Agents PickAI Engineering lesson43 minin the course
- Structured Output Costs Right Answers: One JSON Box, MeasuredAI Engineering lesson25 minin the course
- Week 2
Reliability
395 minDo the maths on success rates, and move decisions out of the model.
- Agent Reliability Math: pass@k, pass^k, and Why 95 Percent FailsAI Engineering lesson80 minin the course
- Deterministic Scaffolding: The LLM Explains, It Does Not ArbitrateAI Engineering lesson80 minin the course
- Durable Execution: An Agent Run Is a Workflow, Not a RequestAI Engineering lesson39 minin the course
- Reflection and Decomposition: Making Agents Check Their Own WorkAI Engineering lesson80 minin the course
- Context Rot: Measuring How Long Context Degrades Your AgentAI Engineering lesson80 minin the course
- Multi-Agent Systems, and When Not to Build OneAI Engineering lesson36 minin the course
- Week 3
The governor: limits, cost and approval
380 minPut hard limits on steps, time and money, and a person before risky steps.
- LLM Agents in Production: The Model in a LoopAI Engineering lesson70 minin the course
- Rate Limits and Provider Failure: The Number One Production ErrorAI Engineering lesson90 minin the course
- Token Cost Engineering: Where the LLM Bill Actually GoesAI Engineering lesson80 minin the course
- Model Routing and Cascades: Running Three Models Without ChaosAI Engineering lesson85 minin the course
- LLM Guardrails and Safety: Wrapping the Model So It Can ShipAI Engineering lesson55 minin the course
- Week 4
Tracing and safety
464 minSee every step the agent took, and stop it from being turned against you.
- LLM Observability and Tracing: Seeing Inside a Nondeterministic SystemAI Engineering lesson75 minin the course
- Prompt Injection and the Lethal Trifecta: Why LLM Agents Leak DataAI Engineering lesson60 minfree
- Defending Against Prompt Injection: The Layered PlaybookAI Engineering lesson60 minin the course
- Tool and Sandbox Security: Locking Down What an Agent Can DoAI Engineering lesson75 minin the course
- Agent Identity and Authorization: Whose Authority Is It?AI Engineering lesson70 minin the course
- MCP in Production: What the Protocol Buys and CostsAI Engineering lesson39 minin the course
- MCP Security: Tool Poisoning, Rug Pulls, and the Confused DeputyAI Engineering lesson85 minin the course
- Milestone
Design a controlled agent
140 minA tool loop, a governor, tracing and human approval, in one design you defend.
milestone: A tool loop, a governor (limits on steps, time and money), tracing, and human approval.
- Design AI Agentinterview guide, free25 minfree
- Design an AI Agent PlatformSystem Design lesson70 minin the course
- Mock interview: design AI AgentAI interviewer, senior level45 min
Questions about this path
- What is a governor in an agent?
- Code outside the model that sets hard limits: how many steps the agent may take, how long it may run and how much it may spend. When a limit is hit, the agent stops, whatever the model wants.
- Why do agents need human approval?
- Some steps cannot be undone, like sending money or deleting data. The agent proposes the step, a person checks it, and only then does it run.
- Do I need the AI Engineer Foundations path first?
- It helps. If tokens, prompts and tool calls are new to you, start there. If you already call models from code, start here.
This path runs through the AI Engineering course. Here is the whole course, at your price.
The course this path is part of
AI Engineering
204 lessons · about 211 hours · 10 free to read
$49 once, lifetime access · pay with PayPal
Also from me
System Design Masterclass
770 lessons · about 283 hours · 18 free to read
$49 once, lifetime access · pay with PayPal