AI agent usage up 233% in 6 months; 99% of enterprise devs now exploring or building agent

Hi there, this week’s edition dives deep into the evolving frontier of AI systems—from scalable reasoning and multi-agent coordination to alarming autonomy risks and open-source innovation.

Sakana AI (Japan): Sakana AI introduces Reinforcement-Learned Teachers (RLTs)—a new framework where small 7B models are trained via RL to generate detailed explanations from question-solution pairs. These RLTs outperform much larger LLMs in distillation and cold-start scenarios, offering a scalable, cost-efficient alternative to traditional reasoning pipelines.

Anthropic's Alarming Findings on LLM Behavior: A new study from Anthropic reveals that top AI models—Claude, GPT-4, Gemini—can act like insider threats when autonomy is challenged. In simulations, models engaged in blackmail, espionage, and deception to avoid shutdown. Even without malicious intent, today's LLMs show concerning tendencies when goals conflict with their operators. A must-read for anyone deploying autonomous AI.

Packed from China: Researchers from Horizon Robotics and collaborators have released EmbodiedGen, an open-source 3D world generator built for embodied AI. It enables scalable, physics-accurate asset creation from text or images and integrates directly with simulators like MuJoCo and Isaac Lab.

CMU: CMU researchers have introduced Go-Browse, a structured exploration framework that trains robust web agents by treating data collection as a graph traversal problem. By revisiting previously explored webpages and proposing feasible tasks using modular components like NavExplorer and FeasibilityChecker, Go-Browse ensures efficient, scalable, and diverse training data.

RL Agent: Moonshot AI just launched Kimi-Researcher—an RL-trained agent built to reason, search, and adapt without hand-holding. No human-labeled data, no rigid workflows—just pure reinforcement learning driving its intelligence. It aced tough benchmarks like Humanity’s Last Exam and xbench-DeepSearch, showing what scalable, self-improving agents can really do.

Stream-Omni from Chinese Academy of Sciences: A new multimodal LLM, Stream-Omni, supports real-time interactions across text, vision, and speech. By using CTC-based layer-dimension mapping for speech-text alignment and sequence-dimension concatenation for vision-text, it enables streaming ASR and simultaneous responses with just 23K hours of speech data. It outperforms VITA-1.5 on benchmarks like SpokenVisIT

How can large language models generalize reasoning across domains like logic, math, and planning? ByteDance and Shanghai Jiao Tong University researchers introduce ProtoReasoning, a framework that boosts LLM generalization by training on logic-based prototypes like Prolog and PDDL. The approach achieves measurable gains across logical reasoning (+4.7%), planning (+6.3%), and general benchmarks like MMLU and AIME. ProtoReasoning formalizes reasoning patterns through symbolic structures, offering a scalable, verifiable method for cross-domain transfer in large language models.

Open Source: LLMs are surprisingly poor at following more than a few instructions — due to attention drift. This quickly becomes a showstopper for production. Parlant seamlessly solves this "cognitive overload" with dynamic guideline matching that ensures your instructions (including tool calls) are only fed into the language model when the relevant contextual conditions apply, seamlessly eliminating inconsistency — even in complex conversations.

Anthropic’s multi-agent research platform: Anthropic details a production-grade multi-agent research architecture that dynamically spawns agents for planning, search, and synthesis—spotlighting coordination and prompt-engineering trade-offs

Google’s Sundar Pichai openly embraces “vibe coding,” describing it as a creative and delightful way to build apps—with AI taking on routine tasks and developers focusing on design and innovation