Agentic Systems & Reasoning (AGENT) Intelligence Hub
Real-time accumulated intelligence, primary research, and executive digests for Agentic Systems & Reasoning (AGENT).
Meta AI uses second agent as memory coach
Meta AI researchers propose pairing an unmodified “action agent” with a separate “memory agent” that reviews recent steps, updates a structured memory bank, and selectively injects brief reminders to keep long-running tasks on track. The approach addresses behavioral state decay, where crucial task information becomes buried or lost in long histories and no longer guides decisions. Unlike simple summarization, the memory agent decides whether stored execution states should influence the next action. In benchmarks such as Terminal-Bench 2.0 and tau2-Bench, selective reminders improved task completion rates over baselines and the Mem0 memory layer. Meta has released code on GitHub and highlights open questions about training and scheduling memory interventions.
Read Original Source ↗Cogent AI releases VR-1 cyber reasoning model
Cogent AI has unveiled VR-1, a frontier cyber reasoning model designed to compose and verify full enterprise attack paths, launched alongside IntrusionBench and the Cogent AI Harness security runtime. Unlike general-purpose models, VR-1 is post-trained specifically for cybersecurity investigations, operating under strict time and turn limits and targeting large enterprises with complex cloud, identity, and infrastructure footprints through a gated access program. Benchmark results show VR-1 proving roughly twice as many attack paths at lower cost than several leading models, prompting interest from high-regulation sectors and increasing expectations for more rigorous, execution-based evaluation of AI security agents.
Read Original Source ↗NVIDIA releases Molt PyTorch agentic RL framework
NVIDIA’s NeMo team has released Molt, an Apache 2.0 licensed, PyTorch-native agentic reinforcement learning framework designed as a compact, roughly 8.6K-line codebase for researchers and AI coding assistants. Molt simplifies rapid algorithm iteration by keeping agents as plain Python programs, integrating Ray for placement, vLLM for rollouts, and NVIDIA AutoModel with FSDP2 for training, while maintaining token-exact trajectories and policy-version consistency. Although framed as research infrastructure rather than a production service, Molt targets labs with multi-node H100 or H200 access, supporting applications like tool-use agents, vision-language environments, LLM-as-judge reward loops, and large-scale on-policy distillation.
Read Original Source ↗