The crewAI team released version 1.15.12, introducing new capabilities for tool usage and project initialization. This update includes a new tool for reading arbitrary URLs and consolidates the scaffolding process under a unified CLI command.
Shieldstral is a 3B open-weights multimodal safety classifier that frames content moderation as a policy-adaptive question-answering task, allowing it to accept plain-language policies at inference time without retraining. It unifies text and image safety evaluation and delivers calibrated safety scores across diverse benchmarks while running efficiently on a single 16GB NVIDIA GPU.
The AutoGPT team released version 0.6.68 of the autogpt-platform-beta, introducing updates to the public platform documentation and improving the product tour experience.
LangGraph version 1.15.4 has been released, primarily focusing on the promotion of the Skills Repository feature from experimental to stable status.
Cactus has post-trained the Gemma 4 E2B model to generate a confidence score between 0 and 1 for every response, enabling developers to route uncertain queries to larger cloud models. The team achieved this by adding a lightweight 68k parameter probe layer that reads intermediate hidden states during decoding to predict the probability of an error.
Researchers present TerraZero, a procedural driving simulator and self-play training stack designed for training robust autonomous driving agents without human demonstrations. The system utilizes a configurable C engine to run simulation on the CPU and policy inference on the GPU over a zero-copy path, sustaining 1.3M agent-steps per second on a single server-grade GPU.
Researchers introduce ProofCouncil, an LLM agent designed to tackle open mathematical problems using an author-critic architecture. The system served as a submission to the second batch of FirstProof, a challenge involving 10 real-world mathematical problems that agents must solve autonomously.
The Mach-Mind-4-Flash Technical Report introduces a 35B-parameter Mixture-of-Experts (MoE) agentic model with only 3B activated parameters. Achieved through post-training optimization without scaling pre-training compute, it performs on par with or surpasses 100B-parameter-class models.