Anthropic is advising companies to consider merging subagents into a single agent that loads context as needed. In one comparison, the single-agent setup ran about twice as fast, and Anthropic has published a migration kit on GitHub.
The New York Times suggests five tasks, such as applying for jobs or booking hard-to-get restaurant tables, where agents like Meta's Muse and OpenAI's dots can prove their worth.
The researchers behind OpenMobile-2 argue smartphone agents improve by training on many different apps rather than on more data. They built an open environment with over 110 apps, including 35 replicas of popular ones, and say their model scores highest among open agents on real Android devices.
Microsoft Research has released Agensh, a multi-agent harness with no central orchestrator in which self-organising workers share a Git repository, a chat and a findings board. On ProgramBench, with a six-hour budget, it scaled from one agent to 1,024.
Cognition has released the Devin Vulnerability Fixes plugin, three MIT-licensed agent skills that inventory, patch and test PHP, Node.js and self-managed GitLab deployments affected by this week's critical remote code execution flaws. The plugin then opens one pull request per repository.
Nanbeige LLM Lab has released Nanbeige4.3-33B, a 33-billion-parameter dense model under Apache 2.0 with a 512K context window. It scores 80.9 on Terminal-Bench 2.1 and will be used to teach the forthcoming Nanbeige4.3-3B.