45 LESSON DAYSBEGINNER → PRODUCTIONPROJECT-FIRST
AI AGENT ENGINEERING
From fresher to
production engineer.
A project-first path for engineering graduates who want to design, build, evaluate, secure, deploy, and operate reliable AI agents across 45 structured daily milestones.
Your progress0 / 45
THE OUTCOME
Not another prompt-engineering course.
You will leave with systems you can run, measure, defend, and explain—not a folder of disconnected notebooks.
- Design controlled agent loops and safe tool contracts
- Build grounded RAG with measurable retrieval quality
- Evaluate behavior before it reaches production
- Deploy with tracing, budgets, security, and runbooks
FULL CURRICULUM
The production path
Seven modules across 45 lesson days. Every chapter pairs a lesson with a lab you can actually run.
01 Days 01–08 · 32 hours
Engineering Foundations
Software, LLM, and Measurement Foundations
+
Engineering Foundations
Software, LLM, and Measurement Foundations
1.1
Day 01: Course Setup, Typed Python and the Production Mindset Today you will build the first version of the AI Service Desk Agent. It is intentionally simple.
1.2 Day 02: Async Python, Event Loops and Concurrent I/O Yesterday the Service Desk application was deliberately simple: receive data, validate it, and return a result.
1.3 Day 03: HTTP Clients, API Failures, Rate Limits and Backoff An AI application rarely lives alone. It calls model providers, vector databases, SaaS APIs, internal microservices, identity systems, and observability backends.
1.4 Day 04: How LLM Applications Work + Structured Output Until now, most of our Service Desk behavior has been deterministic: given the same validated input and the same program state, our code follows rules we wrote.
1.5 Day 05: Evaluation Harness: Stop Building by Vibes Teams often build an AI feature, try five examples manually, decide it “looks good,” and continue adding complexity.
1.6 Day 06: Prompt and Context Engineering with Cost Awareness A prompt is not decoration around an LLM call. It is one of the control surfaces of the application.
1.7 Day 07: FastAPI Fundamentals: Serve the First AI Service So far, much of our Service Desk logic can be called directly from Python. A production application needs a stable boundary that other systems can use.
1.8 Day 08: Streaming with SSE, Cancellation & Partial Failure An LLM may take time to produce a full answer. If an API waits for the entire result before sending anything, the user sees silence and may assume the application is stuck.
02 Days 09–12 · 16 hours
Core Agent Loop
Core Agent Loop, Tools, and Failure Engineering
+
Core Agent Loop
Core Agent Loop, Tools, and Failure Engineering
2.1
Day 09: Tool Calling: Model Requests, Runtime Executes A plain LLM can generate text, but it cannot directly read your production database, reset a password, create a ticket, or query a private API unless the application gives it a…
2.2 Day 10: Build the Agent Loop from Scratch Yesterday the model could request a tool. Today we allow the system to repeat that process until the task is complete.
2.3 Day 11: Failure Engineering: Timeouts, Duplicates and Loop Control Yesterday we built the core loop. Today we assume every dependency can fail and ask what the runtime should do next.
2.4 Day 12: UNASSISTED GATE 1: Agent Loop from a Blank File Today is not about adding another framework or feature. It tests whether the mental model from Days 01–11 is actually yours.
03 Days 13–17 · 20 hours
RAG Infrastructure & Retrieval
RAG Infrastructure, Hybrid Search, and Poisoning Safety
+
RAG Infrastructure & Retrieval
RAG Infrastructure, Hybrid Search, and Poisoning Safety
3.1
Day 13: Embeddings and Semantic Retrieval People rarely describe a problem using the exact words used in documentation.
3.2 Day 14: RAG Ingestion: Parsing, Chunking, Metadata and Storage A common mistake is to blame the vector database when RAG results are poor. Often the real problem was created during ingestion.
3.3 Day 15: RAG Retrieval + First Quantitative RAG Evaluation Yesterday we prepared documents for retrieval. Today we connect retrieval to generation.
3.4 Day 16: Hybrid Search, Filtering and Reranking Semantic search is excellent when wording differs but meaning is similar. Keyword search is excellent when exact identifiers, product names, error codes, or rare terms matter.
3.5 Day 17: Indirect Prompt Injection and RAG Poisoning A RAG system brings external text into the model's context. That text may contain instructions, whether intentionally or accidentally.
04 Days 18–22 · 20 hours
State & Graph Workflows
State, Sessions, Graph Runtimes, and HITL
+
State & Graph Workflows
State, Sessions, Graph Runtimes, and HITL
4.1
Day 18: State, Sessions, Persistence and Memory A model call is temporary. It receives context, produces output, and ends.
4.2 Day 19: Context Compaction, Memory Poisoning and Prompt Caching A long-running agent accumulates messages, tool results, retrieved documents and instructions.
4.3 Day 20: Why a Graph Runtime? LangGraph State, Nodes and Edges We can keep adding if, elif, loops and flags. Eventually the control flow becomes difficult to see because the workflow is hidden inside imperative code.
4.4 Day 21: Persistence, Checkpoints and Fault Recovery A graph execution may run for seconds, minutes, or much longer if humans are involved.
4.5 Day 22: Human-in-the-Loop, Interrupts and Idempotent Side Effects Autonomy is not the objective. Correct, safe task completion is.
05 Days 23–29 · 28 hours
Multi-Agent & MCP Standards
Multi-Agent Architecture, MCP, and Interoperability
+
Multi-Agent & MCP Standards
Multi-Agent Architecture, MCP, and Interoperability
5.1
Day 23: Orchestration Patterns: Router, Planner, Evaluator and Agents-as-Tools As agent systems grow, teams sometimes reach immediately for “multi-agent. ” Often a simpler pattern is enough.
5.2 Day 24: Multi-Agent Systems: Only When One Agent Is Not Enough A multi-agent system contains multiple agent runtimes or agent-like specialists with distinct responsibilities and communication boundaries.
5.3 Day 25: UNASSISTED GATE 2: Timed Broken-Repository Debugging Today is intentionally different from a normal lesson. You are given a broken repository and limited guidance.
5.4 Day 26: MCP 2026-07-28: Architecture and Migration Reality AI applications often need access to external capabilities: ticket systems, databases, files, internal APIs, search services and automation tools.
5.5 Day 27: Build a Local MCP Server with STDIO STDIO transport lets a host start a local MCP server process and communicate over standard input/output streams.
5.6 Day 28: Remote MCP over Streamable HTTP: Auth, Discovery and Caching A local STDIO server inherits many protections from the local machine and process launch.
5.7 Day 29: MCP Security: Tool Poisoning, Supply Chain and Least Privilege MCP makes it easier to connect capabilities. That convenience also makes it easier to connect an unsafe capability if trust is not designed explicitly.
06 Days 30–40 · 44 hours
Evaluation & Production Ops
Production Hardening, Security, Observability, and Ops
+
Evaluation & Production Ops
Production Hardening, Security, Observability, and Ops
6.1
Day 30: Production Streaming API: Backpressure, Disconnects and Progress Events Day 08 introduced streaming as a user-experience and protocol concept.
6.2 Day 31: Identity, Authorization, Multi-Tenancy and PII An agent with tools can read and change real systems. The runtime needs to know who is asking, what they are allowed to do, and which tenant or organization the data belongs to.
6.3 Day 32: Reliability: Retry Budgets, Circuit Breakers, Queues and Backpressure Earlier we learned how an individual HTTP call can retry a transient failure. That is useful at small scale.
6.4 Day 33: Tracing and Reproducing Nondeterministic Bugs An agent request can cross many boundaries: API, model, retrieval, tools, databases and remote services.
6.5 Day 34: Advanced Agent Evaluation: Task, Tool, Trajectory and Judge Quality For a normal text-generation feature, we may care mainly about the quality of the final response.
6.6 Day 35: Regression Engineering: Prompt/Model Versions, Drift and CI Gates A dependency upgrade, prompt edit, model version change or retrieval change can alter behavior without causing a compiler error.
6.7 Day 36: Cost and Performance Engineering LLM price is only one part of total system cost.
6.8 Day 37: Load, Capacity and On-Call Thinking Production engineering asks how the system behaves with many simultaneous users, slow dependencies and traffic spikes.
6.9 Day 38: Docker & Production Packaging A Python application that works on one developer laptop is not yet a deployable production artifact.
6.10 Day 39: OpenAI Agents SDK as a High-Level Contrast We also added termination, state, handoffs, guardrails, tracing and persistence concerns ourselves.
6.11 Day 40: Interoperability and Ecosystem Literacy: A2A + Learning New Frameworks from Docs Yesterday we looked at a high-level agent SDK. Earlier we learned MCP for connecting an AI host to capabilities such as tools and resources.
07 Days 41–45 · 20 hours
Enterprise Production & Capstone
Capstone Project & Portfolio Defense
+
Enterprise Production & Capstone
Capstone Project & Portfolio Defense
7.1
Day 41: Capstone Freeze: Requirements, Architecture and Unseen Acceptance Tests The capstone is where the course stops behaving like a guided tutorial and starts behaving like an engineering delivery.
7.2 Day 42: Independent Implementation and Integration A model client can work in isolation. A retriever can work in isolation.
7.3 Day 43: Adversarial Security and Data-Protection Day A normal functional test asks whether an expected user can complete a task.
7.4 Day 44: Chaos, Load, Eval and Deployment Gate A system can pass functional tests and still fail under dependency outages or concurrency.
7.5 Day 45: Architecture Defense, On-Call Drill and Portfolio Evidence A production engineer must be able to defend why the system is designed the way it is, not only show that it runs.
READINESS
Who this is for
Recent engineering graduates, junior developers, cloud or DevOps engineers, and builders who know basic programming but need a structured path to production AI.
Before you begin
- A laptop that can run Python and Docker
- Basic programming logic
- Willingness to debug and document work
- No prior ML degree required
READY?
Your first agent should not be your first production lesson.
Build the foundations first, then earn the autonomy.
Start Day 01 →