Videos / AI

Episode 13 of 22 · AI, end to end

Agent frameworks: what they add and what they hide

Chapters

1. The loop, said out loud0:00
2. Four moving parts0:17
3. One turn, in full0:33
4. The transcript is the state0:49
5. The stop condition is yours1:05
6. Bolted on, not built in1:22
7. The plumbing under the sink1:38
8. One: wait, then double1:56
9. Two: tokens as they land2:12
10. Three: the repair loop2:29
11. Four: a run is a tree2:45
12. Five: crash on step nine3:01
13. A river and a rail map3:18
14. Control flow as a model output3:34
15. Where the loop wanders3:50
16. States and edges up front4:08
17. Where the map runs out4:25
18. Choose by variance4:43
19. The editor's desk5:00
20. A subagent is just a tool call5:18
21. Handing over the desk5:37
22. What does not cross the seam5:55
23. Three of everything6:12
24. One good agent, usually6:30
25. The bonnet welded shut6:48
26. Half the prompt is not yours7:06
27. Who packs the context7:22
28. One number, no breakdown7:41
29. Two quiet ones more7:56
30. Cut a window in it8:11
31. Buy the tools8:27
32. Four files stay yours8:44
33. And the rest is plumbing9:01
34. The afternoon test9:19
35. The exit test9:35
36. Start smaller than that9:53
37. The whole argument10:08

Agent frameworks: what they add and what they hide. Every agent framework you have ever installed is the same loop you already wrote, with things bolted on. Call the model. Read what it asked for. Run the tool. Put the result back.

Mark as watchedSaved in this browser only.