Multi-agent systems stop sharing one workflow: OOPMAS builds agents and workflows per query, reaching 89.6% on mixed tasks
Synopsis
Automated design of multi-agent systems has mostly worked at the task level, optimizing one fixed workflow per benchmark and applying it to every query. OOPMAS moves design to the individual query: a frozen LLM as meta-agent first generates object-style agents with roles, tools and persistent state, then an executable main-function workflow, while a dynamic skill library accumulates structured lessons from execution feedback without gradient updates. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy, 18.1 percentage points above the strongest baseline, ADAS.
Interpretation
Automated generation of multi-agent systems moves from the task level to the query level: each query gets its own agent set and its own coordination workflow instead of one fixed pipeline shared across a benchmark. Earlier automated design methods optimized a single workflow per benchmark and applied it uniformly to every query in that benchmark, implicitly assuming one fixed structure is optimal for all queries within a task. On a benchmark mixing code, math and QA, OOPMAS reaches 89.6% overall accuracy against 71.5% for the strongest baseline, ADAS, an absolute gap of 18.1 percentage points; on the competition-level Hard Mixed benchmark, OOPMAS reaches 90.8% against 72.3% for ADAS.
The agents themselves are synthesized per query, not merely selected from a fixed pool and rearranged into a workflow. Earlier query-level methods generated a workflow per query but drew agents from a fixed predefined set, so the workflow was limited by the expressivity of those agents. The meta-agent first generates a shared agent pool in one call for a task distribution, then runs a coverage check per query, adding specialist agents and merging them into a query-specific set when a required capability is missing.
Workflows are executable main functions over object-style agent classes and improve across optimization rounds through a dynamic skill library, with no gradient updates, distillation or fine-tuning. The skill library stores structured failure and fix records as in-context optimization memory for the meta-agent, rather than a callable library of executable code. After each round, outcomes are classified as workflow crash, wrong output or success; for the first two the meta-agent produces a root cause, fix summary and fix reasoning, then a corrected workflow is re-executed, and the library is organized into round-level score trajectories, crash patterns by error type, successful root-cause/fix pairs and failed fix attempts.
Perspective
Evaluation covers two benchmarks: a mixed benchmark drawn from HumanEval, MBPP, MATH, GSM8K, HotpotQA and DROP, with samples taken from validation splits, and a competition-level Hard Mixed benchmark drawn from AIME 2025/2026, LiveCodeBench, CodeContests and GPQA Diamond. By default GPT-5 serves as both meta-agent and worker model.
OOPMAS depends on a strong meta-agent: on GPT-5-nano, AFlow's fixed MCTS workflow scores higher than OOPMAS, indicating that online synthesis of agent code is less reliable with a weak backbone. The skill library currently improves workflow generation only within a single optimization run and does not yet persist across deployments, domains or future query streams; per-query generation also raises inference cost substantially, and hard problems require multiple repair attempts. Both benchmarks remain benchmark-based evaluations, and broader deployment may introduce tool-use, long-context, safety, privacy and distribution-shift challenges not fully covered.