Microsoft Research has introduced AgentRx, an automated framework designed to pinpoint the exact moment when an AI agent's trajectory becomes unrecoverable. The system targets failures that emerge in long, probabilistic, multi-agent interactions, where traditional task-completion metrics provide limited insight. AgentRx automates tracing and evidence collection to replace reliance on coarse performance signals. Developers can use it to diagnose failures and improve the reliability and safety of AI systems. Microsoft suggests the tool could support debugging for cloud incidents and complex web interfaces. No public release date or broader availability details were provided.

• Identify the exact failure point in an AI agent's trajectory
• Trace long, probabilistic multi-agent interactions automatically
• Replace coarse task-completion metrics with evidence-based diagnostics
• Support safer, more reliable AI deployments in cloud and web contexts

NVIDIA today introduced Nemotron 3 Super, an open hybrid Mamba-Transformer mixture-of-experts MoE model designed for agentic reasoning. The model is built to support reasoning, coding, and long-context analysis while remaining efficient enough to run continuously at scale. Multi-agent systems can generate up to 15x the tokens of standard chats, including history, tool outputs, and reasoning steps, leading to context explosion. The so-called thinking tax of using massive reasoning models for every sub-task makes multi-agent applications expensive and slow. Nemotron 3 Super is a 120B total-parameter model with 12B active parameters in its MoE organization. NVIDIA positions the model as a step toward scalable, agentic AI capable of sustained reasoning.

• Address context explosion with a 120B total, 12B active-parameter MoE
• Reduce per-task compute via MoE routing
• Scale reasoning for continuous multi-agent contexts
• Enable reasoning across coding, planning, and long-context analysis

Minisforum announced a flagship NAS designed to run large language models locally, with OpenClaw pre-installed. The N5 Max is powered by a Ryzen AI Max+ 395 Strix Halo APU, featuring 16 Zen 5 cores up to 5.1 GHz, a Radeon 8060S iGPU with 40 CUs, an XDNA 2 NPU, and 64MB of L3 cache. Official details on storage capacity and pricing were not released. The unit can be configured with 32GB to 128GB of system memory. Minisforum positions the NAS as a platform for local AI inference on a compact form factor, enabling on-device LLM workloads.

• Enable local LLM deployment on a compact NAS form factor
• Support 32GB–128GB memory configurations
• OpenClaw pre-installed for flexible AI tasks
• Ryzen Strix Halo APU with XDNA NPU for on-device inference