Growing behavior trees as agent harness: an agent that writes its own reflexes

CatAstro_Piyush · x · 2026-11-19

Sihan Xu publishes a 15-minute essay (with code) arguing that future agents won't sit waiting for user turns — they'll act in the world and keep responding to what happens around them. Two core ideas: (1) behavior trees may be the best harness design, letting routine situations be handled without waking the model; (2) an agent can grow its own tree by learning new reactions from lived tasks, so the reflex tree starts at one node and gets denser over time. Framed via Silver & Sutton's 'Era of Experience' argument that agents need long-running experience streams.

Original post →

More from coding & agent

coding & agent channel →