Microsoft’s Sept. 25 Copilot redesign puts Home (Chat, Cowork and Office), Code (solution-building on GitHub Copilot technology) and Autopilot (a persistent proactive agent) in one work surface. The operationally important detail is that long-running Cowork, Code and Autopilot plus frontier models are usage-billed; Agent 365 adds spend policies, approval workflows, model-family controls and usage history, with Copilot Studio cost management planned for October. This is a major Microsoft enterprise-agent distribution and governance move. Source: https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/
NVIDIA announced its Open Agent Safety Platform, pairing broadly available open-source OpenShell (runtime boundaries, action tracing and policy enforcement) with Sentry, a BlueField-4 out-of-band watchdog reference design that NVIDIA says can quarantine boundary-breaking agents in milliseconds. Microsoft is among the announced ecosystem participants. The runtime software can extend to Arm and Intel compute, while Sentry’s hardware enforcement depends on NVIDIA infrastructure; performance and efficacy claims are vendor-reported. Source: https://nvidianews.nvidia.com/news/open-agent-safety-platform
Emerging-model radar: Aleph Alpha released Kolibri-1 on Oct. 3, a German/English reasoning MoE with 78B total but 3.46B active parameters, tool calling, a stated 1M-token context (vendor recommends up to 262K for efficient serving), Apache-2.0 weights and a self-hosted vLLM path. This is a newly surfaced European sovereign/open-weight option for bilingual retrieval and agent workflows; it is not a small-hardware model despite its low active parameter count (FP8 weights require about 78 GB). The model card and specs are provider-reported; Reddit discussion signals early community discovery, not independent benchmark validation. Sources: https://huggingface.co/Aleph-Alpha/Kolibri-1 ; https://www.reddit.com/r/LocalLLaMA/comments/1wwl7y6/alephalphakolibri1_hugging_face_78b_parameters/
Emerging-model radar: Xiaomi released MiMo-V2.6-Distill-Qwen-9B alongside open RL resources, making a purpose-trained agentic 9B checkpoint and reproducible training materials available for local experimentation. Xiaomi reports >7,000 RL task environments spanning software engineering, vulnerability reproduction, knowledge-intensive work and web development, and gains when RL-training from the distill baseline (for example SWE-bench Verified 61.1→66.2 and Terminal Bench 2.1 37.1→52.8). These are vendor-reported benchmarks; early LocalLLaMA users have also reported looping/tool-use instability, so the gains should not be read as settled independent performance. Sources: https://mimo.mi.com/docs/en-US/news/latest/v2-6 ; https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B ; https://www.reddit.com/r/LocalLLaMA/comments/1wrq71o/mimo_v26_flash_mopd/
New evaluation work: AgBench provides open artifacts for reproducibly comparing local, hybrid and cloud execution of agents on personal AI devices. Its paper reports 162M+ measured data points and finds no universal winner: local-only runs avoid API costs and cloud data exposure but generally lose task success and latency, especially under concurrency; hybrid trade-offs depend on what is sent to the cloud. This offers a practical deployment/evaluation lens beyond model leaderboard scores; results are paper-reported. Source: https://arxiv.org/abs/2609.38652