What This Is
An experiment by the open-source research team lmgame shows: 12 small AI Agents working with a central reasoning model achieved a 96% case-solving rate, while a group of small Agents self-governing through voting only reached 38%. This number is worth pausing for any company currently scrambling to deploy "multi-Agent" systems.
They built a detective game in a virtual Victorian London: 100 addresses, 10 suspects, find the killer within 12 hours. They tested four configurations — 12 small Agents gathering evidence with a large model in command, a rule-script in command with a small Agent swarm, a large model going door-to-door solo, and a small Agent swarm self-governing through voting. The results: large model + small Agent swarm scored 96 (standard case), 93 (case with planted false evidence); large model solo: 57; small Agent swarm solo: 38. The largest gap appeared in the "someone is lying" case — the large model would press on the origin of the evidence, while the rule-based system merely counted votes.
Industry View
This research pours cold water on the current "more Agents = smarter" narrative. Its core conclusion: small Agents are good for parallel search (divide), large models are good for integrated judgment (conquer) — the division of labor is where value lives. Anthropic has also publicly acknowledged that roughly 90% of enterprise Agent projects stall on engineering execution, not on model capability.
But we should also see the boundaries of this experiment: it's a controlled game scenario. In real business, evidence is messier, chains are longer, and error tolerance is lower. "12 small Agents" may look lightweight, but the reasoning API calls underneath aren't cheap — this division-of-labor architecture demands serious engineering orchestration that most enterprises simply don't have. Also, the rule-based system matched the large model on the standard case; that doesn't mean rules are better — it just means many scenarios don't need the most expensive option. That's a cost-saving signal, but also a ceiling signal on capability.
Impact on Regular People
For enterprise IT: Before deploying Agents, map out clearly whether each step is "evidence gathering" or "judgment." Don't treat the large model as a cure-all, and don't pile on multi-Agent just because it's trendy.
For individual careers: Those products claiming "a swarm of AIs working for you" will most likely underperform a single thinking AI plus a few tool plugins. Don't pay extra for the "multi-Agent" label.
For the consumer market: A wave of "multi-Agent collaboration" C-end apps will surface this year. We recommend watching for three months, checking real user reviews, before paying — don't rush the launch.