Script to Swarm
Stop watching agents.
Start shipping them.
A text-first course on autonomous AI agent engineering. From a single Python asyncio script to multi-agent swarms running in production with evals, traces, guardrails, human-in-the-loop, and a public portfolio.
Twelve stages. Seven capstones. Every pattern an AI agent developer actually ships.
If any of this sounds familiar
You don’t need another “agents” tutorial.
You need to ship one.
- You’ve watched fifteen LangGraph videos. You couldn’t ship one of them next Monday.
- Your “agent” works on the happy path and loops forever on the second prompt.
- You hand-rolled a tool calling loop and now have no idea what broke when it stops calling tools.
- You added “memory” by stuffing the whole transcript into the context window.
- You have zero evals, zero traces, and no idea how often your agent quietly hallucinates.
- Your portfolio is screenshots. Recruiters cannot tell if you can actually build one.
Most people stay stuck watching tutorials. Builders get hired. This course is the path from one to the other — twelve stages, seven capstones, a public portfolio at the end.
Seven capstones
You graduate with seven shippable agents.
Every capstone is a self-contained agent you build, deploy, and put on your GitHub. Each one targets the patterns from the stage that precedes it. By Capstone 7, your repo reads like a senior AI engineer’s.
Capstone 1 · After Stage 3
Toolbot
A structured-output tool-calling agent that researches and reasons over web sources with Pydantic-validated outputs and recoverable tool errors.
Capstone 2 · After Stage 4
Memo
A memory-backed personal assistant. Short-term buffer + vector recall + scheduled compression + cross-session sync. Knows you across sessions.
Capstone 3 · After Stage 5
Scholar
A ReAct-based deep-research agent. Bounded iteration, self-reflection, graceful degradation, citations the user can audit.
Capstone 4 · After Stage 6
Review Crew
A multi-agent code-review swarm: planner, reviewer, security auditor, summarizer. Supervisor pattern with explicit handoffs.
Capstone 5 · After Stage 8
Eval Harness
An automated eval harness for any agent. LLM-as-a-judge, regression suite, hallucination metrics, CI integration.
Capstone 6 · After Stage 10
Hardened Support Agent
A production-ready customer support agent. Prompt-injection defenses, PII redaction, sandboxed execution, audit log, compliance gate.
Capstone 7 · Stage 12
Portfolio Drop
An open-source agent of your choosing, deployed, documented, demoed, and contributed back to a library. The piece recruiters actually click.
Syllabus
13 chapters · 55 units
Every unit is principle → do → verify → reference. Click any free unit to read it now. Paid units unlock after purchase.
Ch 0Start herePsychological2 units· ~15 min
What this is. Who it's for. The honest test for whether you're ready to build agents instead of watching them.
Ch 1Python + Async FoundationsMedium5 units· ~55 min
Agents are concurrent by nature. If you can't reason about asyncio, FastAPI, and event-driven flow, every agent you ship will be a race condition waiting to happen.
Ch 2LLM Fundamentals for AgentsPEAK5 units· ~53 min
The model is not the agent. The model is one expensive, fallible function inside an agent. Treat it like one.
Ch 3Tool Calling + Structured OutputsPEAK5 units· ~90 min
An agent without tools is a chatbot. Tools are how the model touches the real world — typed, validated, recoverable. (Capstone 1.)
Ch 4Memory + State ManagementPEAK5 units· ~94 min
An agent without memory is a goldfish with an API key. Decide what to remember, where to store it, and when to forget. (Capstone 2.)
Ch 5Single Agent WorkflowsPEAK4 units· ~84 min
ReAct, plan-and-execute, reflection. One agent, one loop, bounded — and able to fail gracefully. (Capstone 3.)
Ch 6Multi-Agent OrchestrationPEAK5 units· ~105 min
Most multi-agent systems are unnecessary. The ones that aren't need supervisors, message contracts, and handoff discipline. (Capstone 4.)
Ch 7Human-in-the-Loop SystemsMedium4 units· ~40 min
The cheapest way to make an agent safe is to let a human review the dangerous parts. Build the seam before you need it.
Ch 8Evaluation + Quality AssurancePEAK4 units· ~82 min
If you can't measure the agent, you can't improve it — and you definitely can't ship it. (Capstone 5.)
Ch 9Observability + TracingMedium4 units· ~41 min
Agents fail invisibly. Traces, spans, costs, and alerts are how you make the invisible obvious.
Ch 10Security + GuardrailsPEAK4 units· ~92 min
Prompt injection is the SQL injection of this decade. Plan for the hostile user before the friendly one. (Capstone 6.)
Ch 11Production DeploymentHigh point4 units· ~45 min
An agent on your laptop is a demo. An agent on Kubernetes with canaries and rollback is a product.
Ch 12Open Source + PortfolioIndependent4 units· ~89 min
Builders get hired because their work is visible. Ship the agent, write the architecture, record the demo. (Capstone 7.)
Full access
Unlock all 12 stages.
- ✓All 13 chapters, 55 units, every future update
- ✓Seven capstone projects you keep, ship, and put on GitHub
- ✓Chapter 0 + the first unit of Chapter 1 are free — taste the format first
- ✓Reference repos for every capstone, MIT-licensed, fork friendly
Frequent objections
Honest answers.
Who is this for?
Engineers who can write Python at a working level and want to actually build autonomous agents that run in production. If you've never written async code or hit an API, start with Blank to Browser or Prompt to Production first. If you've built a chatbot demo and felt like you were faking the rest, you're the target reader.
Why Python? Can I use TypeScript?
The course is Python-first because the agent ecosystem — LangGraph, CrewAI, vLLM, SGLang, the major eval and tracing libs — is still Python-native. Every pattern translates to TypeScript; the SDK calls swap one-for-one. But the worked code is Python.
Will this go out of date when LangGraph changes APIs next month?
The principles won't. The patterns — ReAct, supervisor, message passing, human-in-the-loop, eval harness, sandbox — are the durable layer. The illustrative SDK code is explicitly disposable and quarantined. If LangGraph changes its API tomorrow, you'll still know what an agent loop is and why bounded iteration matters.
How long does the whole course take?
About 35–50 hours of reading and doing, plus another 20–30 hours across the seven capstones if you actually ship them. Most learners take 6–10 weeks at an hour or two per evening. The capstones are designed so you can pace them however your life allows.
Do I need an OpenAI / Anthropic key?
Yes for the capstones. You can use any provider that supports tool calling — Anthropic, OpenAI, Google, or a local model via vLLM (covered in Stage 11). Most capstones run for cents in API costs. The course never assumes a specific vendor.
What does my portfolio look like at the end?
Seven public GitHub repos, each with a deployed agent, a README that explains the architecture decisions, and a short demo. Stage 12 walks you through writing the architecture docs and recording the demos so they're actually persuasive to a hiring manager.
What if I get stuck on a capstone?
Every capstone has a reference repo, an architecture diagram, and a known-bug list. The 'verify' beat of every unit tells you exactly what success looks like — so when you're stuck, you have a named failure mode to grep, ask your AI for, or compare against the reference.
Most people stay stuck watching tutorials.
Builders get hired.
Chapter 0 is free. Read it in under fifteen minutes — by the end you’ll know whether to keep going.