Videos / AI

Episode 12 of 22 · AI, end to end

Agents from first principles: build the loop yourself

Chapters

1. One call, or many0:00
2. Text in, text out, full stop0:16
3. Letting the text name an act0:32
4. Three moving parts0:49
5. The bigger prompt trap1:06
6. The definition, tight1:24
7. A stack of cards1:41
8. The loop, in full1:56
9. The reply has two shapes2:11
10. Your code runs it, not the model2:28
11. Append, never rewrite2:44
12. The token goes round3:02
13. A menu3:18
14. Name, schema, sentence3:35
15. The schema is not a guard3:50
16. The description is the interface4:06
17. Few tools, sharp edges4:23
18. What a tool returns is a prompt4:40
19. The meter runs4:56
20. The turn limit, crude5:13
21. The model says it is done5:30
22. A terminal tool5:46
23. Cycles, and budgets6:02
24. Decide what cut off says6:19
25. Circle it, or slam it6:36
26. What an exception costs6:52
27. Catch, and return7:08
28. Write the failure for a reader7:23
29. Watch it recover7:39
30. Two limits on forgiveness7:56
31. A rented engine8:14
32. Four things you own8:29
33. Context is a policy8:46
34. Guardrails, per tool9:03
35. A failing run is a transcript9:20
36. The tasks you score it on9:37
37. The whole loop9:54

Agents from first principles: build the loop yourself. A pocket calculator answers once and then stops. A person holding one keeps going: press, read, press again, until the number is right. That second thing is an agent. Not a smarter

Mark as watchedSaved in this browser only.