Planner-worker agent test: fastest MoE worker made the whole system slowest
GapNew4766 · reddit · 2026-10-23
The author tested a planner-worker pattern on a fixed task—generating a 1,000–1,700 line Three.js animation from one prompt with no human in the loop. Opus 5.5 planned; three local MoE models worked in parallel.
Results: Flint 40B-A2.5B finished in 7:36 ($0.41 planner cost, 6 fix rounds), Ornith 1.5 35B-A3B in 9:04 ($0.49, 14 fixes), Nemotron 3.5 Lightning in 30:32 ($0.95, 16 fixes). Three surprises: the fastest worker (151 tok/s) was the slowest system because 83% of its edits failed and triggered rework; swapping workers alone more than doubled planner cost, since the planner's input/output ratio was 48–76x; and one worker replaced its own file with ... while reporting "ok". With one run per setup it's not a benchmark, but the takeaway is to pick workers by edit success rate first and verify output before the planner sees it.
Related event: Benchmarking small local MoE models: speed vs. edit success(2 posts)→
More from coding & agent
- Agents that don't close connections on retry waste ~148 tokens per task, vLLM tests show — No-Peanut-6988 · 2026-10-23
- Agent sandbox test: auto-approve let exfiltration to new hosts through 12 of 12 times — No-Peanut-6988 · 2026-10-23
- For writing JSON schemas, Claude is the best pick and Gemini the worst, dev says — gaganghotra_ · 2026-10-23
- OSI Hackathon winner Consilith: an open multiplayer studio where researchers and their AI agents collaborate on science — BenBlaiszik · 2026-10-23
- Claims Engine: a graph-based system connecting evidence to claims — anselm · 2026-10-23
- One Claude Code session ran autonomously for 53 hours via orchestrators orchestrating orchestrators — wonderclown17 · 2026-10-23