Agent performance in production hinges on infrastructure, not the agent

Nvidia argues AI security needs defined requirements, enforceable controls, named owners and evidence protections work, applied across models, harnesses and runtime environments. Example: an agent reading malicious instructions in an attached document should be blocked by network policy, with protected logs capturing the tool call and destination.

Unitree's Dex5-S robotic hand packs 22 degrees of freedom and costs $6,500 per hand, per a Reddit post in r/Singularity. The Wuji Hand 2, a 20-DoF alternative, is priced between $16,000 and $20,000.

Openlayer's jevals swaps free-form LLM-as-judge scoring for typed Jev decisions in evals. Posted as a Show HN project on GitHub.

Hugging Face shipped tokenizers v1, a major version of its text tokenization library, with the release post focused on encode, decode, and scaling measurements.

V7 turns scattered company files into context agents can use for complex, source-linked work, built on GPT-5.6.