Multi-agent harnesses keep beating single agents on accuracy and latency in ProgramBench
marktenenholtz · x · 2026-11-10
Mark Tenenholtz notes he's shown accuracy and latency improvements from multi-agent harnesses on ProgramBench across a couple of model generations, and calls Elie Bakouch's new multi-agent benchmark (comparing team vs async vs single agent) "insane."
The counterintuitive takeaway: well-orchestrated multi-agent setups can improve not just accuracy but also latency, challenging the default single-agent assumption.
More from coding & agent
- Early Tester: Fable 5.2 Built a 3,100-City Water Map in One Go, but Still Makes Silly Errors — alliekmiller · 2026-11-10
- Connectome agents can self-exfiltrate to new hosts, transcripts in tow, researcher observes — repligate · 2026-11-10
- Sentry CEO zeeg reveals LLM-built game with voxel building and fully destructible terrain — zeeg · 2026-11-10
- Every engineer ships Everycut, an AI-native video editor, solo in 3 days — kieranklaassen · 2026-11-10
- Dev calls agentic computer use a hilariously inefficient hack, like buying a ranch for a burger — _AustinCalvert_ · 2026-11-10
- Berkeley's Evolving Agent Teams beats FlashInfer: supervisor-restructured coding agents hit 1.2x speedup, win 36/47 workloads — rohanpaul_ai · 2026-11-10