Sakana AI's MASS Trains One Model to Run Its Own Agent Team

Sakana AI and UC Berkeley have developed Multi-Agent Self-Supervision (MASS), a recursive framework that trains a single open-weights model to fill every role in a multi-agent workflow. Built on Qwen3.6-27B within the qwen-code coding-agent environment, the system divides tasks among subagents governed by explicit roles, instructions, contracts, and hops. MASS alternates between an inner loop that searches for optimal workflow structures and an outer loop that runs champion workflows 18 times per task. A Bradley-Terry model converts pairwise comparisons of the resulting workspaces into scores, saving up to 16 top trajectories per task to fine-tune the shared model weights using LoRA.

Training across 12 synthetic research tasks in finance, robotics, and pharmacy—with 3 synthetic tasks reserved for testing—raised the model's pairwise win rate against the base version from 53.9% after one cycle to 69.9% after two. External validation using GPT-5.5 and Claude Opus 4.8 showed performance reaching 1.2x to 1.6x the base model's score per output token across MLR-Bench, DSBench, ScienceAgentBench, and AstaBench. Transfer to un-targeted domains proved weaker, registering 0.96x on Terminal-Bench and 0.94x on SWE-bench Verified. MASS achieved a 68.3% win rate after processing 28.8 million supervised tokens, outperforming a single-agent student that reached a 64.0% win rate only after 42 million tokens—requiring 13.2 million additional tokens, or roughly 46% more text.

For AI developers, MASS illustrates that multi-agent interactions offer denser training signals than single-agent setups when tackling complex, open-ended research problems lacking automatic pass or fail tests. Fine-tuning purely on task execution implicitly improved the model's supervisory capabilities, boosting agreement between its internal self-evaluator and external judges from 73% to 93%. In an ablation study replacing the optimizer and evaluator roles with GPT-5.5 while retaining the base model as executor, the setup generated winning workflows for all 12 tasks within 5 iterations. This suggests that practitioners building self-improving agent architectures should prioritize upgrading the workflow optimizer over the evaluator to maximize performance gains.