Sakana AI has introduced Multi-Agent Self-Supervision, or MASS, the company said in a post on X and a blog on 11 October 2026. In MASS, one shared model works as a team of subagents, searches for a better team workflow, judges the results itself and is then trained on its best team’s work. The work was a collaboration with researchers at UC Berkeley, and the paper was posted to arXiv on 8 October.
Starting from the open-weights Qwen3.6-27B, MASS models won 53.9% of comparisons against the base version on three held-out synthetic research tasks after one round and 69.9% after two, as judged by GPT-5.5 and Claude Opus 4.8, the blog says. On four public research benchmarks, score per output token reached 1.2 to 1.6 times the base model’s level.
Business Pill · VERIFICATION COST
A one-minute explainer of verification cost: work that is cheap to make is not cheap to check. It teaches the general idea only and says nothing about any person or organisation in this story.
The key insight: As we read it, the key design choice is who grades the work. MASS lets the model judge its own team’s output and trains on what it picks, and Sakana measures progress with outside judges kept away from the loop, so the safeguard is the measurement, not the training signal.
The blog describes a single model filling three roles: the agent that solves a task, the workflow optimizer that revises how subagents divide the work, and the evaluator that compares two finished workspaces and picks the better one. Sakana calls this “homogeneous recursive self-improvement”.
In an inner loop, the model searches for a better team workflow with its weights fixed. In an outer loop, it runs the best workflow 18 times per task, ranks the results itself, keeps up to 16 trajectories per task and fine-tunes on them with LoRA. The workflow text is removed from the training prompts, so the model learns to organise the work from the task itself.
The training tasks are non-verifiable, which the blog defines as tasks where “an automatic rule-based signal is not available”. GPT-5.5 and Claude Opus 4.8 assess the finished work separately, and their judgments are withheld from MASS.
External-judge win rates on Sakana AI’s three held-out synthetic research tasks after one and two rounds of MASS on Qwen3.6-27B, as reported in Table 1 of Sakana’s MASS blog; the judges were GPT-5.5 and Claude Opus 4.8.
What Was Posted on X
Introducing Multi-Agent Self-Supervision (MASS): Recursive self-improvement through collective intelligence
As AI systems tackle increasingly open-ended research problems, human experts may struggle to assess results and guide further progress. The model itself may be the best available optimizer and evaluator. The challenge is obtaining useful self-supervision within that homogeneous loop, where a model risks reinforcing its own blind spots rather than correcting them.
We ask whether a model can learn from the collective intelligence of its own team.
We introduce Multi-Agent Self-Supervision (MASS). One shared model solves tasks through a team of virtual subagents, searches for better multi-agent workflows, then uses its own judgments to select the best ones.
Training on the best team’s executions provides self-supervision by distilling the team’s collective experience into the shared model.
We ran two cycles using a 27B open-weights model on synthetic open-ended research tasks. After the second cycle, score per output token reached 1.2-1.6x the base model’s level across four research benchmarks.
Three key results:
First, MASS enables role generalization within an RSI loop: training only on task-solving data from the optimizee naturally improves the capabilities of both the optimizer and the evaluator.
Second, multi-agent trajectories are more efficient training data than single-agent trajectories for RSI.
Third, optimizer capacity, rather than evaluator capacity, can be the bottleneck in the speed of improvement.
We believe the community needs to investigate homogeneous RSI carefully to support AI safety. When a model is also its own evaluator, shared blind spots could allow mistakes to be accepted as progress and reinforced through training. For AI to improve beyond human expertise safely, we need to understand these failures and preserve effective ways to detect them and intervene.
Sakana AI’s post of 11 October 2026 introducing Multi-Agent Self-Supervision (MASS), with a link to the blog this piece reads in full. The results in the piece are Sakana’s own.
What the Rounds Produced
The synthetic suite has 12 tasks in finance, robotics and pharmacy; nine are for training and three are held out. On the training tasks, the win rate against the base model was 65.6% after one round and 82.5% after two; the second-round model also won 60.8% of comparisons against the first-round model on the test tasks.
On public benchmarks the gains were uneven. Score per output token rose on MLR-Bench, DSBench, ScienceAgentBench and AstaBench, but was 0.96 times the base value on Terminal-Bench 2.0 and 0.94 times on SWE-bench Verified, the blog says. Sakana suggests adding software engineering tasks to the training data.
Supervised training tokens processed by two students trained from the same base model, and their win rates against it, as reported in Sakana AI’s MASS blog.
Three Findings Sakana Draws
Sakana says training only on task-solving data also improved the other two roles: its agreement with the external judges rose from 73% to 93% across the cycles, and the base model running first-round workflows won 64% of comparisons against the base model running its own.
Sakana says multi-agent trajectories can give more per training token. The first-round MASS model reached a 68.3% win rate against the base after 28.8 million supervised training tokens, against 64.0% after 42 million for a model trained on single-agent trajectories; Sakana puts that at 31% fewer tokens.
Sakana says the bottleneck can be the optimizer rather than the evaluator. With GPT-5.5 as both evaluator and optimizer, winning workflows were found for all 12 tasks within five iterations; upgrading only the evaluator helped much less, and the fully homogeneous loop reached 11 of 12 tasks more slowly.
The Structural Read
As we read it, MASS moves effort from the model to the workflow first: the inner loop changes only how subagents divide work, exchange results and check one another, before any weights change.
The outer loop then folds that workflow back into the weights, since the workflow text is removed from the orchestrator’s training prompts so that, as the blog puts it, the model learns to organise the work from the task itself.
Sakana itself flags the risk of the setup: in its words, “When a model is also its own evaluator, shared blind spots could allow mistakes to be accepted as progress and reinforced through training.”
Sakana AI, post on X introducing MASS, 11 October 2026
“We ask whether a model can learn from the collective intelligence of its own team.”
Three Implications
TEAMS BUILDING AGENTS Sakana reports that the base model running workflows found by the first-round model won 64% of comparisons against the base model running its own.
TRAINING BUDGETS The blog reports a 68.3% win rate after 28.8 million supervised tokens of multi-agent data, against 64.0% after 42 million of single-agent data.
SAFETY AND EVALUATION With a stronger external model as both evaluator and optimizer, winning workflows were found for all 12 tasks within five iterations; the self-only loop reached 11 of 12.
The framework puts it this way: “The surface, not the model, governs autonomy”.
As we read it, MASS tests a version of that claim inside the lab: the gains start in the workflow around a fixed model, and only then are trained into the model itself.
What Is Not Established
The results come from Sakana’s own runs on a synthetic suite whose task descriptions were generated by GPT-5.5, with external models as judges. The blog reports two rounds on one 27B model; it does not show how far the loop goes, and it names shared blind spots, where a model accepts its own mistakes as progress, as an open safety question.
Business Engineer Framework
The Agentic Harness War
A Business Engineer framework on why the harness around a model, not the model alone, now decides what agents can do.
Sakana AI’s MASS trains one model on the self-judged best work of a team made of copies of itself. After two rounds on Qwen3.6-27B, the blog reports a 69.9% win rate against the base on held-out tasks and 1.2 to 1.6 times its score per output token on four research benchmarks, with lower scores on two coding benchmarks.
A note on sourcing. On 12 October 2026 we read Sakana AI’s post on X of 11 October (through a mirror of X’s public data), its blog Multi-Agent Self-Supervision in full, and the abstract page of the paper on arXiv (2610.12176). Every figure here comes from the blog; the results are Sakana’s own and were not independently reproduced by us. Nothing here is a forecast.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
