Researchers propose MUFFIN, a system that lets AI agents work asynchronously by posting intermediate findings to a shared memory buffer, while keeping each agent’s private reasoning local. An orchestrator routes questions and sends harder ones to multiple agents.
The authors report that MUFFIN had the highest accuracy in 17 of 20 benchmark settings. Against multi-agent debate, they report a 19.2-point accuracy gain while using 77.4% fewer tokens. The abstract does not name the benchmarks or provide independent validation, so the results’ wider relevance is unclear.
Why it matters
For teams building multi-agent AI systems, MUFFIN suggests agents may exchange useful findings with fewer tokens, but the abstract does not identify the tests or establish how well the approach works beyond them.
