AI Struggles with Complex Tasks, Google-MIT Study Reveals

The numbers don't lie. For every 1,000 tokens spent, a single agent solved 67 tasks. A coordinated multi-agent team finished 21.

Context gets lost between handoffs. Isolated errors in committees then multiply—catastrophically. Without shared information, mistakes compound up to 17 times faster.

A central coordinator only reduced that error factor to four. The study pinpoints a breaking point: once a solo agent succeeds 45% of the time, adding more provides no benefit. The coordination tax consumes every gain.

So the promise is real but brittle. It demands a job that can be perfectly partitioned, a rare condition. For the messy, sequential work that defines real problems, a solitary agent isn't just simpler.

It's fundamentally more competent. The industry's rush to orchestrate committees might just be adding overhead where none is needed. Sometimes, the most advanced system is the one that works alone.