Scale Labs open-sources AgentEnv, a modular multi-agent reinforcement-learning framework that runs the same task across Claude Code, OpenAI, and custom agents using Environment and Agent Cards, with 48 built-in step types and four sandbox options.
Living Models says its Gemma 4–powered BOTANIC-1 system can pinpoint causal crop mutations in hours, ranking the validated CmEIN3 melon flower-sex mutation first among 2,494 candidates while running locally on one NVIDIA L4 GPU.
OpenAI says an unreleased AI model solves the Navier-Stokes Millennium Prize problem using 10,000 concurrent agents and creates an independent mathematician advisory group to vet that claim and scores of other reported results.
EverMind AI launches Raven, a pre-alpha self-evolving multi-agent harness that builds DAG workflows for complex tasks, shipping with Research, Code, Design and Oncall agents plus 13 third-party presets through ACP, CLI and OpenAI-compatible APIs.
Datalab releases Omni Extract Bench, an Apache-2.0 open-source benchmark that evaluates structured-document extraction from Datalab, Reducto, Extend and LlamaExtract with normalized JSON scoring, Hungarian array matching and provider-paid API runs.
arXiv imposes a universal two-paper-per-month submission limit starting October 1, 2026, responding to a surge in AI-written, fragmented manuscripts that helped drive 40,363 submissions and nearly 9,000 support tickets in September.
Fig launches RIDGE, finding no clear frontier-AI champion as Astra, Opus 5.5 and rivals show jagged task-level results; Astra leads VisualWebArena at 139 of 177 tasks, yet every lower-scoring model solves some tasks it misses.
Anthropic says 950 Claude agents make a molecular-biology discovery in 21 hours by identifying an uncatalogued repeating pattern around a known enzyme, but University of Copenhagen biologist Mario Rodríguez Mestre says his team had already found it and raises concerns about whether Claude learned from their chats, which Anthropic denies.
Anthropic adds build-eval and hillclimb commands to Claude Code’s API skill, enabling automated evaluation creation and iterative app optimization with held-out tests; on a 44-ticket support benchmark, the workflow reaches 98.9% accuracy at about 1 cent per ticket.
Cambridge researchers warn that automating AI research and development could compress years of progress into months, risking loss of control over superhuman systems as AI already writes most code at the companies building it.