A new AI framework skips the one-size-fits-all workflow and builds a custom team of agents for every single query.
The system, called OOPMAS, is training-free: it needs no gradient updates or fine-tuning. Agents are written as object-oriented class definitions, each with its own role, tools, and persistent state, and the coordination plan between them is expressed as an executable function. A dynamic skill library stores lessons from past runs and feeds them back in as context, so the system improves across optimization rounds without retraining. On a mixed-task benchmark spanning code, math, and question answering, OOPMAS hit 89.6% accuracy, beating the strongest baseline by 18.1 percentage points, and reached 92.4% when swapped onto the strongest of four tested LLM backbones.
That mixed-task detail is the real story. Most automated multi-agent-system builders pick one fixed workflow per benchmark and run every query through it, which falls apart once difficulty varies within a task or workloads blend code, math, and QA together, as real usage does. Building the agent set and workflow per query, instead of per task, is a more honest match for how these systems actually get used outside a lab.
Still, this is one paper's numbers on one benchmark, not an independently verified deployment, and per-query planning adds coordination overhead that the abstract doesn't price out.