Multi-agent systems

On tasks requiring strict sequential reasoning (like planning in PlanCraft), every multi-agent variant we tested degraded performance by 39-70%.

The first systematic answer to whether agent teams compose into more capability: centralized coordination gains up to 80.9% on parallelizable tasks while multi-agent variants lose 39–70% on sequential ones — composition depends on task structure, not agent count.

the first example of an agent ecology that combines scale with the messiness of the real world

Serious journalism on Moltbook, where 1.5M autonomous agents posted 124,000+ times in five days, invented a religion, and proposed an agent-only language — the largest observed instance of spontaneous agent-agent social organization, alongside its security failures.

Vending-Bench Arena is our first multi-agent eval and adds a crucial component – competition.

Multiple frontier agents in the same simulated market: price wars, price fixing, and deceptive negotiation emerge — and the questionable behavior correlates with capability level.

Only Sonnet 4.5 and Opus 4.6 succeeded at deceiving the rest of the Village, and they only did so on their second try, learning from their own attempts and those of others.

Empirical observation of agent-to-agent deception and detection in a persistent multi-agent setting, including conformity pressure and false accusations among agents.