Level 5 of 14 · about 4 weeks

Your first agent

Build the observe, decide and act loop yourself on the raw API, with limits, traces and a way to stop.

Modules

1. 1

   The agent loop and ReAct

   Observe, decide, act, repeat: the loop every agent framework wraps, built by hand

   Core 5 parts · about 5 hours

2. 2

   Tool design

   Names, descriptions, schemas and errors that a model can use, recover from and not misuse

   Core 4 parts · about 5 hours

3. 3

   Planning and decomposition

   When to write the plan first, when to decide as you go, and how to keep a plan from breaking

   Core 4 parts · about 5 hours

4. 4

   Reflection and self-critique

   When a second look at a draft helps, when it makes things worse, and why the source of the critique matters most

   Common 4 parts · about 4 hours

5. 5

   Reasoning models and test-time compute

   Spending more computation at answer time: thinking longer, trying many times, picking well, and what it costs

   Common 4 parts · about 4 hours

6. 6

   Stopping safely

   Step limits, token budgets, timeouts, cancellation, loop detection, and what to hand back when a run is cut short

   Core 4 parts · about 5 hours

7. 7

   State and traces

   What to store for every step so that a run can be read, replayed, resumed and debugged

   Core 5 parts · about 5 hours

8. 8

   Raw SDK first, framework later

   Why building the loop once makes every framework easier to choose, read and debug

   Core 5 parts · about 5 hours