Multi-agent orchestrator dilemma: cheap models rubber-stamp failures, big models burn cash
Muted_Ad_9442 · reddit · 2026-09-03
A developer shares a multi-agent setup where four roles (planner, validator, worker, mechanical) each run a chain of models with fallbacks for rate limits, tiered by how much judgment the role needs.
The orchestrator problem: it does both clerical work (dispatch tasks, write back results) and judgment work (spot wrong outputs, distinguish real failures from noise, keep parallel tasks from colliding).
- A top reasoning model in that seat was noticeably expensive even on mechanical tasks
- A small model agreed to everything: a sub-agent returned success having done nothing, and it recorded the task as complete
Current compromise: split the role — cheap model for dispatch and bookkeeping, big model only for actual judgment calls like "is this finding real, fix or just log it". Log analysis showed those calls are roughly 1 in 10 actions but cause nearly all the damage when wrong.
The author asks how others configure orchestrators and how to stop small models from accepting bad work.
More from coding & agent
- Vector Ingestion at 50M Rows: The Pitfalls Your 1,000-Doc Prototype Won't Reveal — victorialslocum · 2026-09-03
- Reddit user swears by local AI harness Vellum: free, proactive, but nobody talks about it — Cooperman411 · 2026-09-03
- Qwen Code v0.23.0 ships Anthropic stream hang fix, daemon memory tasks and more — qwen-code-ci-bot · 2026-09-03
- Dev Builds an Interactive 'Multiverse TV' Where Twitch Chat Picks What Plays Next — d_pit · 2026-09-03
- Brian Kernighan and Rob Pike wrote a regex matcher in a few dozen lines of C — blaizedsouza · 2026-09-03
- tinyfool: doing research with AI is like having grad students write your code — tinyfool · 2026-09-03