Researchers have found that AI agents solve math problems more reliably when the thinking and the tool use are handled by separate, smaller agents instead of one model trying to do both.
The framework, called MSARL, splits the work into two roles. A Reasoning Agent breaks down the problem and decides which tools to call. Separate Tool Agents handle the actual tool operations, like running code, and each is trained specifically for that job using a mix of imitation learning and reinforcement learning with rewards tailored to its role. Tested on math problems that require code execution, the setup produced more stable reasoning and more accurate final answers than single-agent systems that try to juggle both jobs at once. The researchers also found the approach held up across other tool-use tasks, not just math.
This matters because most tool-using AI systems today still ask one model to plan and execute in the same breath, which is a bit like asking someone to do long division while also operating the calculator. MSARL's bet is that cognitive-load interference, not raw model size, is what breaks down in long multi-step tasks. If that holds up outside the paper's test cases, it is a cheaper fix than training ever-larger single models to be better multitaskers.
It is also a bet against the industry's default instinct to throw a bigger model at coordination problems. Smaller, specialized agents dividing labor is a less flashy pitch than one giant model doing everything, but it is the kind of unglamorous engineering that tends to survive contact with production.