Week 13 · Season 4 · Will
It decides the next step
You can build an agent loop, choose when to use a workflow, set step, time and money limits, explain the SDK, and show each step in a live feed.
This week
- An agent is a loop: tools, stop reasons, a step limit
- Workflow or agent: the simplest design that passes
- The $400 night: budgets, headroom, runaway loops
- What an agent SDK hides: turns, budgets, gates, hooks
- Show every step: events and a live feed
How you learn
Before you write the loop, be it: take a calendar and a to-do list, and plan a week one step at a time, writing down what you looked at, what you decided and why, and where you would stop.
What you build
FRIDAY plans your week in several steps: a loop with three tools, step and time limits checked before every call, an honest stop reason, and a live feed of plain lines, proved with a pretend model.
Stretch: Give FRIDAY a Draft-and-check pass: a second call judges the plan against your done lines, with a two-round limit and a best-plan fallback.
What you ship
Project 4: The morning briefing. Every morning FRIDAY gathers your calendar, tasks and notes and briefs you, by itself, within budget.
- WEEK-PLAN.md and a loop diagram
- agent.py, tools.py, feed.py and tests with a pretend model that pass in CI
Lessons
- An agent is a loop
After this deck you can draw the agent loop from memory (ask the model, run the tool it asks for, send the result back, repeat), read a stop reason, and name the four ways a loop should end.
- Workflow or agent?
After this deck you can tell a workflow (you write the path) from an agent (the model picks it), and choose the simplest design that passes your tests at a price you can afford.
- The $400 night
After this deck you can read a runaway agent's log, find why it never stopped, and give a loop three limits (steps, seconds and dollars) checked before every call, with headroom.
- What the SDK hides
After this deck you can say what an agent SDK runs for you (the loop, built-in tools, permission checks, hooks, squeezing old history), set turn and budget limits, and read the result it hands back.
- Show every step
After this deck you can turn an agent's events (a tool starts, a tool returns, the run ends) into a live feed people can read and trust, with honest failures, a stop button and the final cost.
- Go deeper: routers, helpers and judges
After this deck you can name and sketch three agent patterns (routing, orchestrator-workers and evaluator-optimizer), and give each one a stop rule and a fallback.
Labs
Python in the browser, next to the lessons.
Weekend build
- FRIDAY, week 13: plan my week
FRIDAY decides its own next step: it looks at your week, thinks, acts, looks again, and shows every step.
Skills you prove
- Raw agent
- Build a tool-using agent with an MCP server in an afternoon, without a framework
Free links for this week
Hand-picked official docs, papers and talks.