OpenAI Enables Autonomous Agent Management with 'Symphony'
OpenAI has published the new Symphony specification. Whereas developers previously had to manage multiple Codex sessions by hand, Symphony lets AI agents pull tickets directly from Linear and execute them autonomously until the task is complete.
As OpenAI points out, "human attention is the bottleneck." An architecture in which agents pull and execute tasks on their own could fundamentally change AI coding workflows. It marks a shift from the traditional model of humans directing agents to one where agents operate autonomously and humans supervise.
The move also aligns with a trend in the Hacker News community, where agent management tools have been launching one after another, including Ruflo (multi-agent orchestration for Claude Code) and Daintree (a delegation environment for AI coding agents). The orchestration layer for AI agents is maturing rapidly.
SoftBank Plans US Listing for New AI and Robotics Company
SoftBank Group is reportedly planning to list a new company focused on AI and robotics in the United States. According to the Financial Times, SoftBank intends to establish a company that combines its years of accumulated robotics operations — including its experience with Aldebaran Robotics, the company behind Pepper — with its recent focus on AI investments.
A US listing appears aimed not only at raising capital but also at strengthening SoftBank's presence in the global AI race. SoftBank is already known as a major investor in OpenAI and other leading AI companies, and this move suggests an acceleration of its AI infrastructure investment.
Chinese Hospitals Sell Anonymized Patient Data to AI Companies
According to Nikkei Asia (via Caixin), multiple hospitals in China have been found selling anonymized patient data to AI companies. As medical data becomes essential for training AI models, the practice is raising ethical and legal concerns.
The data is said to have personal information such as names stripped out, but no clear standards have been disclosed regarding the degree of anonymization or whether patient consent was obtained. Amid China's rapidly growing AI industry, data privacy issues are drawing international attention.
Multi-Agent Web Search Sets a New SOTA — Web2BigTable
Featured in Hugging Face's Daily Papers, "Web2BigTable" is a multi-agent framework specialized for internet-scale information retrieval and extraction. It adopts a two-tier orchestrator-worker architecture and combines an execution-verification-reflection loop with persistent external memory.
It achieved an Avg@4 success rate of 38.50 on the WideSearch benchmark, a dramatic 7.5x improvement over the previous SOTA. It also recorded 73.0% accuracy on XBench-DeepSearch. A distinctive design feature is the sharing of workspaces among agents, which cuts down on redundant exploration.
Medical Data Fed to AI Is Often Incomplete — University of Würzburg Study
According to a paper published in Nature Health by a research team at the University of Würzburg in Germany, the medical information provided to AI is incomplete in many cases. As AI applications in healthcare expand, input data quality is drawing attention as an issue that directly affects diagnostic accuracy.
The study shows that patient information fed into AI systems in real clinical settings frequently contains gaps, undermining the reliability of AI recommendations. The finding that the integrity of the input pipeline matters as much as model performance for practical AI in healthcare carries valuable real-world implications.
A Unified Theory of Decision Trees and Diffusion Models
Also featured in Hugging Face Daily Papers, "Trees to Flows and Back" proposes a theoretical framework that unifies decision trees and diffusion models. It shows that decision trees correspond to diffusion processes in an appropriate limit and derives a common objective function called Global Trajectory Score Matching (GTSM).
Under this framework, gradient boosting is shown to be asymptotically optimal. On the practical side, the paper also proposes two methods: TreeFlow, which speeds up tabular data generation by roughly 2x, and DSMTree, which distills decision tree structure into neural networks. It is a fascinating study that blurs the boundary between traditional machine learning and deep learning.
---