Week 1 · Lesson 03

Agent loop

ReAct lite: plan → act → observe → stop. This is cognition — not persistence yet.

Time · 1 day
Learn
  • Tool registry and observation normalizer
  • Stop conditions: success, error, max steps, timeout
  • Why budgets are part of architecture, not niceties
Build
  1. In-process loop with two tools: content_planning, finish(draft)
  2. Hard limits: MAX_STEPS=8, TIMEOUT_MS=60000
  3. Print every step to the console
Checkpoint

You can watch the model choose tools. Killing the process mid-run corrupts nothing external (there is no world API yet).

Four notes
  • What problem does this layer solve?
  • What is allowed to call it?
  • What must it never do?
  • How do I observe and test it?