Leaked study: Agent wiring swings performance from -70% to +80%
brucemacv · x · 2026-08-14
An X post cites a leaked Google Research and MIT study: holding prompts, tools, and compute constant, only changing agent wiring caused performance to swing from 70% worse to 80.8% better than a single agent, averaging 0%. The author advises testing solo agent first; above 45% success, multi-agent may not help. Also mentions Anthropic's leaked Graph Engineering, claiming 96% quality at 46% cost.
More from AGI Musings
- Does AI remember more mean better? A test with 21 drafts — sujingshen · 2026-08-14
- AI for Introspection: Externalize Experience, But Real-Life Observation Is Key — DevToD4 · 2026-08-14
- In the AI Era, the Scarcest Resource Is Judgment, Not Data — ingliguori · 2026-08-14
- The AI Race Dilemma: Even Rational Actors Lead to Irrational Outcomes — ventblanc · 2026-08-14
- Matthew McDonagh: Use AI for the Right Stuff — McDonaghMatthew · 2026-08-14
- Stewart Alsop III: Six Years of Five Minutes a Day to Get to the Useful Stuff — StewartalsopIII · 2026-08-14