The internet ran on open protocols and open source. So will agentic AI.Help us build them.

Help us build

them

DAILY AGENTIC AI LINKEDIN NEWSLETTER

• Moonshot AI Open-Sources AgentENV for Scale RL Agent Training

  Moonshot AI and kvcache-ai open-sourced AgentENV, a high-throughput distributed system engineered for agentic reinforcement learning. Developed to support RL training for Kimi models, the platform features sub-second environment snapshotting, state resumption, and tree-forking capabilities across thousands of parallel agent interaction loops. Standardizing distributed environment scaling simplifies reinforcement learning infrastructure for developers training complex reasoning agents.

  X

• Microsoft Builds a Cybersecurity Team Out of Agents

  Microsoft launched Project Perception, a system in which red-team agents find attack paths, blue-team agents investigate risk, and green-team agents apply fixes. Its MDASH harness routes roughly 90% of tasks to the specialized MAI-Cyber-1-Flash model and reserves GPT-5.4 for the hardest 10%, scoring 95.95% on CyberGym at half the cost of Microsoft’s previous configuration. The frontier in cyber defense is shifting from agents that find vulnerabilities to coordinated systems that can find, prioritize, and repair them continuously.

  Microsoft

• NVIDIA Gives Engineering Agents the Tools of Physics

  NVIDIA expanded its Agent Toolkit with agent-ready PhysicsNeMo and CUDA-X libraries, while Siemens, Cadence, and Synopsys are integrating the stack into semiconductor and industrial-design workflows. Agents can now invoke sparse solvers, quantum-chemistry calculations, deterministic EDA engines, and the open Nemotron 3 Ultra model; Siemens says its implementation cut library-characterization time by more than 10× while reducing token costs by 5–10×. Following on the heels of BioNeMo release, this one means agents’ work is continuously checked against the same simulation and verification tools engineers trust, “industrializing” agentic AI.

  NVIDIA