---
title: 'In the News: October 12, 2026 (Morning)'
description: 'Sakana AI and UC Berkeley report a model improving its own multi-agent research workflows, and find the optimizer role limits progress more than the judge.'
canonical_url: 'https://darkfactory.dev/news/2026-10-12-morning'
markdown_url: 'https://darkfactory.dev/news/2026-10-12-morning.md'
collection: news
date_published: '2026-10-12T07:25:00-04:00'
date_modified: '2026-10-12T07:25:00-04:00'
---

# In the News: October 12, 2026 (Morning)


One item this cycle. A Sakana AI and UC Berkeley paper loops one model through the roles of worker, workflow designer and judge, and reports what limited the loop.

## 1. Sakana and Berkeley let one model design, run and judge its own agent teams, and the weak link was the designer

**[Recursive self-improvement through collective intelligence (MASS)](https://pub.sakana.ai/mass/)** · Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, Yujin Tang · Sakana AI and UC Berkeley · October 2026

MASS uses one model as the task solver, the workflow optimizer and the evaluator. In an inner loop the optimizer rewrites a written team workflow (roles, instructions, output "contracts" and call order) while the weights stay fixed. In an outer loop the champion workflow runs 18 times per task, the model ranks the results, up to 16 trajectories are kept, and the model is fine-tuned with LoRA on them with the workflow text removed from the prompt. The authors ran two rounds from Qwen3.6-27B on 12 synthetic research tasks in finance, robotics and pharmacy. On the three held-out tasks, judged by GPT-5.5 and Claude Opus 4.8 with their verdicts withheld from training, the win rate against the base model rose from 53.9% after one round to 69.9% after two.

On public research benchmarks, score per output token reached 1.2 to 1.6 times the base model's. On coding it did not: 0.94 times on SWE-bench Verified and 0.96 times on Terminal-Bench 2.0. The authors say they suspect adding software engineering tasks to the training data would fix this. They did not test it.

The ablation most relevant to harness builders: as the optimizer iterated, task-specific information moved out of roles and instructions and into contracts and call order. One evolved finance contract required a data specialist to return explicit checks that no future data leaked into the training split. Swapping in GPT-5.5 for the evaluator alone helped only a little. Swapping it in for both evaluator and optimizer found winning workflows for all 12 tasks within five iterations, and the fully homogeneous loop reached 11 of 12 more slowly. The authors write that "the bottleneck in RSI can be the optimizer, not the evaluator." They also report that workflow search is noisy, with strong iterations sometimes followed by sharp regressions. The post was at 115 points and 70 comments on Hacker News about twelve hours after submission.

**Why it matters:** This is a small study by its authors on synthetic tasks, with three held-out tasks that workflow search also saw, and I did not run the released code. For builders the useful findings are narrower than the RSI framing: the part of a multi-agent workflow that improved was what each agent must establish and hand on, which is a verification contract, and a weak workflow designer capped the results even with a strong judge. It is not evidence that coding agents are improving themselves, since the coding benchmarks went slightly down.