Burkov mocks 2026 agent research: building agents nobody bothers to define

burkov · x · 2026-09-15

ML researcher Andrey Burkov contrasts eras: before LLMs, scientists would 'define an agent clearly and unambiguously first, then propose an algorithm to build one.' In 2026: 'This is our agent that we don't define, but you should assume it has something to do with LLMs and something they call a harness.'

A one-liner that captures the terminological drift and conceptual looseness of current agent research.

Original post →

More from Fun

Fun channel →