Everything it takes to build
a reliable agent.

No videos, no slides — every layer runs for real in your browser. Ask her what this place is.

A real agent — it'll take you wherever you want to start.

your guidefree · no signup
Anatomy of an agent

An agent isn't one thing. It's five parts.

Miss one and it flakes in a way you can't explain. Pick a part below: what it is, a live proof running right here, and the course that teaches you to build it for real.

Functions your agent can call — a calculator, a shipping-cost lookup, an API. The model can't run anything itself; it emits a call and your code answers.

the courses

Write a Python function, give it a name and a description, and expose it as a tool the model will actually reach for instead of guessing.

You hand the model a calculator. Who actually runs it?

model emitscalculate("12.5 * 48")
your runtime runs600
model continues"That's $600 total."

The model never computes anything. It asks; your code answers. Miss this and every agent bug looks like magic.

Build & share
new

Every lesson is a notebook.
So is the agent you build in it.

A doc it has to search, tools that touch real records, a face on the front — each one is a block you stack. Start from the support bot, pull it apart, and see what breaks; or start blank and stack your own. Then share it with a link: whoever opens it gets the same working agent, not a screenshot of one.

free · share with a link
the support bot · 4 blocksstacking
docsearch_policyNorthwind returns policy · chunked + embedded in your browser
toollookup_orderdef lookup_order(order_id) — real Python, the model never sees it
toolraise_ticketdef raise_ticket(order_id, reason) — when the policy needs a human
runsystem prompt"Never answer from memory — always search_policy first."
face · Wren

"Thanks for calling Northwind — I'm Wren. What can I help you with?"

No framework, no deploy. Stack the blocks, hit preview, and talk to it — then delete the "don't remember it" line and watch it start making things up.

Put it together

The Autofill Agent

Not a layer — an application. A tiny page-agent staples three ideas together: your page flattened to numbered text (what it sees), a memory store (what it knows), and a model turning the two into keystrokes. Then poison its memory and watch it hallucinate a policy number — or build your own form and set the trap yourself.

live · ~25 min · your memory grows hands

Nobody touches the keyboard. Watch it fill itself.

Full name agent
Work email
Dietary needs
Register

Your memory store grew hands: it reads the page as text and types. Zero keystrokes from you.

Meet Stitch

The agent at the top only knows what you build.

That was Stitch, and she follows you the whole way — she quizzes you on a lesson you just ran, co-authors in the build space, and points the way everywhere else. What she can actually reach is the five layers above, and you build every one of them. In the Curator you name your own agents and hand each a different slice.

  1. 01

    Build the pieces

    Every course leaves something real behind — a Python function becomes a tool, a query becomes a table, your docs become retrieval, facts about you become memory, a SKILL.md becomes a routine.

  2. 02

    Compose your own agents

    Stitch is the house agent, but she's not alone. In the Curator you name your own: hand one your docs, another your tables, and scope each with skills that decide which sources it may touch.

  3. 03

    Every answer shows receipts

    No black box. Under each reply: the skill that fired, the sources it read, the files it cited by name, and a token ledger of what entered context versus what stayed on disk.

one turn with Stitchglass-box
Draft my launch post — what did we decide on pricing?
doc-lookupfired → retrieval memoryunlocked
search_docs("launch pricing")
→ 2 passages launch-plan.md pricing-notes.md
user_memory
→ "audience: data engineers" · "prefers weekday launches"
Keep launch free — your notes hold the paywall test for September. Post Tuesday morning and lead with the live demo; your crowd is data engineers.
sources launch-plan.md pricing-notes.mdmemory ×2
wired to your stack
ToolsTablesRetrievalMemorySkills
Sharpen the fundamentals

The muscles under all of it.

Every layer above is Python calling tools and SQL hitting tables. When a chapter exposes a gap in yours, sharpen it here — same live runtime, real errors, real output.

For data folks

One entity, five messy sources — run the whole pipeline.

A different beast from the journey above: describe your own entity, get AI-generated dirty data, then map, match, merge, enrich and mint durable IDs on real Postgres — until you can just ask your golden table questions in plain English.