Using a large model as tech lead: speculative-decoding-style agent orchestration

prajdabre · x · 2026-09-26

The author shares his favorite agent pattern: a larger model orchestrates, dispatches sub-agents for subtasks, verifies their output, and ties up loose ends. He likens it to speculative decoding — a cheap sub-agent "drafts" an action and the expensive main agent verifies it.

In his experience the main model is a good "tech lead": if sub-agents mess up, it says "fine, I'll do it myself." Letting one big model do everything — task, verification, goal-checking — goes wrong more often. He also mentions an acquaintance simulating a 3-tier org where they act as VP with director and tech-lead layers beneath.

Original post →

More from coding & agent

coding & agent channel →