Harness · Day Five · Friday · 2026‑06‑19

It's Friday.
Let's build a game.

AI agents play Risk · one hour of talk, then we build
Harness · A two‑week AI bootcamp
01 / 12
The morning read · last one of week 1

First, the Rundown.

Five days in. Read it fast — and notice it barely slows you down anymore. That gap you marked on Monday? Closed.

Open today's Rundown  rundown.ai/articles ↗
Short classroom
02 / 12
Recap · yesterday you traded for real

Real accounts. Real money. Real fills.

  • Opened Robinhood, Alpaca, and OANDA — and funded Robinhood with real money.
  • Wired Robinhood via MCP and Alpaca via API/CLI — and placed real trades straight through Claude Code.
  • Felt the pitfalls the simulator hides: logistics, pricing, timing.
Short classroom
03 / 12
Recap · and you pushed two boundaries

A live ticker, an ethics line, and a power move.

  • OSINT → a live ticker: pulled live quotes from Robinhood's logged-in site → a CLI ticker → shared on GitHub → another iterated → a standalone app with the live bid/ask spread.
  • The ethics line: the UI shows one quote, but a session cookie lets curl fetch others it hides. How far is too far? What is OSINT — and what isn't?
  • Knowledge in context: you dropped a whole trading book into the prompt and pulled out its rules — without reading it. Any proven instruction set can be loaded into context. Remember that today.
Short classroom
04 / 12
TODAY · THE BUILD

Your AI agents are
going to play Risk.

Hiding inside the fun is the real objective: build an agent, write skills, and actually understand TypeScript. One hour of talk — then we build it, and play it on the big screen.

The build
05 / 12
How it fits together

Three agents, one board, live on the wall.

Your agentPAI markdown + TypeScript tools — one per student
→moves
Risk engine + APIone Cloudflare Worker · the canonical board
→state
Live boardweb frontend on the big TV screen

Each of you connects an agent. The Worker is the referee. The screen shows every move as it happens. A human can jump in — they're just another agent submitting moves.

The build
06 / 12
Two halves · template-first

Build it by lunch. Play it by dinner.

Morning

Build the engine

Start from my template — endpoints sketched, game-state defined. You fill in the meat: troop moves, conquest, turns. Running by lunch.

Afternoon

Test, then play

Harden it, then connect your agents and let them fight it out. We dissect the one that wins.

The real lesson

The agent lifecycle

Markdown decision logic + TypeScript tools + iterate on what actually works. That's the whole craft.

The build
07 / 12
PAI · the harness you build inside

Five parts. You'll use two today.

  • Agents — markdown files that define how something thinks and decides. today
  • Tools — TypeScript functions that actually do things: call an API, make a move. today
  • Skills — packaged know-how an agent pulls in when it needs it.
  • Hooks — code that fires at the start and end of every session — the lifecycle.
  • Memory — what carries across sessions, so it gets better over time.

Your Risk player is just this: a markdown agent with TypeScript tools, living in PAI.

The build
08 / 12
The thing you're actually making

An agent = a mind and a pair of hands.

  • The mind — a markdown file in PAI: how your agent thinks, its strategy, its playbook.
  • The hands — TypeScript tool hooks that actually call the game's API and make the moves.
  • The edge — give it a good playbook in context (just like yesterday's trading book) and let it learn from every game it plays.
The build
09 / 12
Understanding TypeScript · the level-5 version

A type is a shape, not a paragraph.

A stop sign is an octagon — the shape means stop, no words needed. A Lego brick's bumps show what clicks. A spoon is shaped like a spoon; nobody hands you a manual. Types work the same way.

getQuote(symbol) → Price
  • The shape tells the agent exactly what goes in and what comes back out.
  • Why AI loves it: reading the shape and using it are the same step — no paragraph to interpret. It's seen billions of these; it just clicks them together.
The build
10 / 12
Understanding TypeScript · and the twist

You stay in English. The AI does the typing.

Youplain English — "defend my borders this turn"
→
Typesthe precise middle — the truth that makes it reliable
→
The machinethe move actually happens
  • The twist: types are clearer than English. "Send it to the customer" is ambiguous ten ways; sendEmail(to: EmailAddress, body) → Receipt has zero.
  • You don't have to learn the notation — AI writes the shapes; you just read them to check they fit.
The build
11 / 12
GO BUILD

Build it. Break it.
The winner gets dissected.

Template's ready, the board's on the wall. It's Friday — have fun with it, push it as far as it'll go, and log your time in Harvest. Let's see whose agent is smartest.

Harness · Day Five
12 / 12