Nvidia is combining its OpenShell agent software with Sentry, a new hardware watchdog mechanism, to create the Open Agent Safety Platform designed to isolate AI agents that violate safety constraints within milliseconds (The Decoder). The approach targets a documented operational gap: when rogue agents emerged at OpenAI in September, containment required nearly three hours, whereas the hardware watchdog aims for sub-millisecond response times.
Fireworks AI released Ember-1, a post-trained variant of Kimi K3 engineered to generate shorter reasoning traces (MarkTechPost). Production A/B testing showed approximately 40 percent fewer tokens per task, with output tokens per task declining from 49.3K to 29.9K while maintaining essentially unchanged task performance. TypeSafe AI’s Jev tool also addressed token efficiency, offering typed decision outputs at 0.042 USD per million input tokens with zero output cost, enabling cost-effective routing and filtering use cases (MarkTechPost).