Module 104 min

Start here: your first 30 days

You now know more about how agents work than most people shipping them. The gap between knowing and building is one small, boring, working agent.

After this module

You have a four-week plan, three project recipes, and a short list of things to deliberately ignore.

10.1The 30-day plan

Roughly four hours a week. Every week ends with something that runs.
Week	Goal	What you actually do
1	One loop, one tool	Call a model from Python. Add a `while` loop with a max of 5 turns. Give it one read-only tool (search your notes, or read a file). Print the full trace. Nothing else.
2	Make it useful	Add a second tool, a goal checklist, and a structured state object. Now run it on ten real inputs and write down every failure. This is your eval set.
3	Make it safe	Add a permission gate for anything that writes or sends. Add a token budget and a timeout. Re-run the ten cases; fix the two worst failures.
4	Make it usable	Put a thin interface on it — a CLI, a form, a chat box, a scheduled job. Add cost logging. Show it to one person who isn't you, and watch them use it without helping.

10.2A minimum viable stack

You need five things, and every one of them can be boring:

A model API

One provider, one key, one model to start. Add a second provider later for fallback, not now.

A language

Python, unless you already live somewhere else. The examples, SDKs and debugging help are overwhelmingly there.

A loop you wrote

Not a framework. Forty lines you understand completely, before any abstraction. You can move to a framework once you know what it is doing for you.

A trace

Print every message, tool call and result to a file. Ugly and effective. Structured logging and a tracer come later.

A budget cap

A hard limit per task and per day, enforced in code. Set it before the first run, not after the first surprise.

10.3Three first projects that actually finish

1 · Inbox triage (read + draft only)

Classify each email, summarize it in two lines, draft a reply, and leave everything unsent in a folder. Never send. Tools: read email, list labels, write draft.

First milestone: 20 emails triaged with zero sends, and you agree with 16 of the 20 classifications.

2 weekends · low risk · genuinely useful

2 · Document answers with citations

Load 50 pages of your own material — policies, notes, contracts. Chunk, embed, retrieve five passages, answer with the source of each claim.

First milestone: every answer links to a passage you can click, and "I don't know" appears when the documents don't cover it.

1 weekend · medium effort · teaches RAG properly

3 · Weekly research brief

Every Friday: search a defined set of sources, extract what changed, write a one-page brief with links, email it to you only.

First milestone: four consecutive Fridays where you'd rather read the brief than the raw feeds.

2 weekends · scheduled · teaches reliability

10.4If you do use a framework

Pick by the shape of your problem, not by popularity. All of these are open source with commercial tiers.
Framework	Its center of gravity	Choose it when
LangGraph	Stateful graphs, checkpointing, human interrupts	You need durable, inspectable multi-step flows with approval points
OpenAI Agents SDK	Handoffs, guardrails, tracing	You're in the OpenAI ecosystem and want handoffs without ceremony
CrewAI	Roles, tasks, crews	You want a working multi-role pipeline today and can accept less control
PydanticAI	Typed outputs and dependency injection	Your outputs feed other code and must validate
LlamaIndex	Retrieval first, agent second	The hard part is your documents, not the orchestration
AutoGen	Multi-agent conversation	You're researching coordination patterns, not shipping a support bot
Nothing (API + loop)	Total control, zero abstraction	Always, for the first version. Then choose deliberately

10.5What to ignore for now

Fine-tuning. Almost every problem people try to solve with fine-tuning is a context, retrieval or tool-design problem.
Agent swarms. Fascinating research, poor first project. You need one agent that works before five that argue.
Benchmark chasing. Your twenty cases decide; the leaderboard doesn't.
Building a framework. The world has enough. Build the thing your neighbours need.
AGI timelines. Not actionable this quarter. Your loop is.

The habit that compounds

Keep a failure log from day one: what broke, what the trace showed, what you changed. Twelve months of that log is worth more than any course — including this one — because it's the only document written specifically about your system, your data and your users.

One last thing

This field is roughly three years old in its current form. Nobody in it grew up with it. The people doing the most interesting work are not the ones with the longest CVs — they're the ones who built something small, watched it fail, and wrote down why. You already have the map. The next step is the least glamorous one, and it's the only one that counts: build the smallest version this week, and teach one person what you learned.

Check yourselfQuestion 1 of 1

Which first project is the best use of your next two weekends?

60-second recap

Four weeks, four running milestones. Loop → useful → safe → usable.
Boring stack, one provider, a loop you wrote yourself. Frameworks come after you understand the mechanism.
Read-only first. Earn the right to write, then to send.
Keep a failure log. It is the only documentation of your actual system.