Agent Loop
Loop After Loop: Smarter With Every Turn
Switching off and forgetting isn’t learning. Each loop, an Agent writes its experience into files, saves anything it did right three times as a skill, and checks before it acts. One song, on how it gets smarter with use.
Watch
Video won't play? Download this episode
If the video won't play inside WeChat, tap ··· at the top right and choose “Open in Browser”.
What this one covers
Start with the shape of the loop. An Agent doesn’t work question-answer, question-answer. It reads the task first (perceive), makes a plan (think), actually calls its tools and does the work (act), then looks back to check the result (feedback). Four steps, one loop. The smarts aren’t in any single answer. They’re in the count.
Next, memory. An Agent with no memory meets you fresh at every boot: the pit it fell into yesterday, it falls into again today, unchanged. Write the experience down as files, load them first next time, and pick up where it stopped — that’s all “smarter with use” means.
Skills are saved up. Do the same thing right three times, and that path gets written down as a skill and shelved in the library. Next time it shows up, no thinking from zero — call it straight in. The fuller the library, the faster the hands.
Last: verify first, then trust. A loop with no feedback is just spinning in place. Every loop ends with a check — right, keep it; wrong, fix it and run again. The number in the song isn’t made up: improve 1% a day for 365 loops and 1.01 to the 365th power comes out to about 37.8. The curve shooting up at the end is drawing exactly that.
Sources and notes for this episode
This is the first episode of the series, covering the foundation concept: the Agent Loop. To be plain about sources: no single outside article served as its basis. “Perceive, think, act, get feedback” is the standard way people describe an agent’s loop, and this site’s topic list carries the entry; the lyrics, the analogies, and this commentary are all written by this site.
The number in the song gets its own note. Improve 1% a day for 365 days and 1.01 to the 365th power comes to about 37.78 — that is, you end up at roughly 37.8 times where you started (not 37.8 times more added on top). It is a compound-growth example under a stated assumption, showing how small steps, repeated, pull a gap open. Real people do not improve at a steady 1% a day, so don’t take it as a promise.