Sakana AI introduced MASS on October 12, a research method in which the same language model solves tasks, improves workflows and judges results. It aims to turn an agent team’s work into useful training material for their shared model. This is not a new consumer product or proof that a system can improve indefinitely without external oversight.
The experiment started with Qwen3.6-27B, an open-weights model with 27 billion parameters. Researchers ran two training cycles. After the second, score per output token reached 1.2 to 1.6 times the baseline across four research benchmarks. This metric combines outcome quality and output volume; it does not mean runtime fell by the same ratio or that the system is 1.6 times smarter in every field.
How a team teaches its shared model
First, the model explores ways to divide tasks among agents, exchange information and check outputs. Weights remain fixed during this search. The evaluator compares results and explains which is better, while the optimizer adjusts the workflow accordingly. Organization can therefore improve before further training changes the model itself.
Next, the system learns from selected execution records: conversations, tool calls and results produced during teamwork. The updated model fills all three roles in the next cycle. This differs from asking a chatbot to check itself: it combines systematic workflow search with weight updates, rather than simply rephrasing an answer within a conversation.
The interesting result and its limits
The synthetic tasks covered finance, robotics and pharmacy. External judges based on other models assessed the outputs, but their scores were not fed into the learning loop. This separation tests whether improvement is visible beyond the system’s self-assessment. However, another model’s evaluation does not automatically establish scientific truth or replace appropriate experiments in the field.
Transfer was also uneven. Score-per-token gains appeared on MLR-Bench, DSBench, ScienceAgentBench and AstaBench. On Terminal-Bench 2.0 and SWE-bench Verified, the ratios were instead 0.96 and 0.94 relative to baseline. Training therefore did not improve every task type. Teams building agents should evaluate their target tasks separately, rather than assume better research results mean stronger performance on all coding work.
The central risk follows from the same feature enabling the experiment: solver and judge share a model, and may share mistakes and blind spots. Training on a highly rated error could reinforce it. MASS offers a practical way to investigate this loop. Its current value lies in the research method and released code, alongside continued independent testing, rather than a promise that an agent can already supervise its own progress alone.