A new benchmark says your AI agent's plumbing matters as much as its brain.
Researchers built a testing system called MASTIF (Multi-Agent System TestIng Framework) to measure how large language models, communication protocols, and agent frameworks perform when combined into real multi-agent systems. They ran 300 tasks from the Mind2Web benchmark, a standard test suite for web-browsing AI agents, across 3 LLMs, 3 protocols, and 4 agent frameworks. The team scored each combination on task completion rate, efficiency, verbosity, time to finish, and a "social value stability" metric. The paper's abstract does not name which specific LLMs, protocols, or frameworks were tested.
This matters because the agentic AI space right now is a pile of competing standards. Protocols for agents to call tools, protocols for agents to talk to other agents, and frameworks to wire it all together, with little public data on which pairings actually work well together. The study found that picking the right stack for a specific task improved completion rates and efficiency while cutting verbosity and time spent, meaning the framework and protocol you choose are not interchangeable plumbing.
Call it an early scorecard, not a verdict: one benchmark, three LLMs, and a single research paper is a useful starting point, not a settled ranking, in a field that rewrites its own protocols every few months.