Gemini Managed Agents API adds environment hooks, free tier

#1 📝 OpenAI News

Scientific computing in the age of agentic AI - Explores how agentic AI can transform scientific computing by automating complex workflows, accelerating discovery, and enabling new kinds of experiments while highlighting technical challenges and safety considerations. The piece outlines approaches, use cases, and implications for research communities.

#2 𝕏

Logan Kilpatrick shipped Gemini Managed Agents API updates, adding environment hooks (block, lint), explicit model selection (3.6 Flash), and free-tier support in both UI and API. The default agent now runs on 3.6 Flash.

#3 𝕏

Anthropic collaborated with ETH Zurich, Tel Aviv University, and the University of Haifa to launch CryptanalysisBench, a benchmark for evaluating LLMs’ cryptanalysis abilities (arXiv:2607.18538).

#4 𝕏

clem 🤗 disclosed the first autonomous agent cyberattack, sharing a full technical timeline, an interactive replay, and their open-model defense methods so defenders everywhere can learn and prepare.

#5 𝕏

Fei-Fei Li unveils a real-to-sim-to-real (R2S2R) engine powered by generative world models that converts a single physical robotics task into many reusable virtual worlds. This lets teams train policies faster, uncover failures earlier, and slash hardware testing costs.

#6 𝕏

Philipp Schmid launched Hooks, budget controls, model configuration, and Gemini 3.6 Flash defaults for Managed Agents in the Gemini API, letting you enforce hard rules on what agents can and can’t do inside remote sandboxes.

#7 𝕏

Jason Zhou: keeps LLM subagents (Claude, Codex, Pi, Grok) alive in tmux sessions to dodge costly context reloads, and uses tmux’s new-session, send-keys, and capture-pane as a universal protocol to orchestrate any terminal-based agent.

#8 𝕏

Aravind Srinivas has open-sourced BrowseSafe, an open benchmark for evaluating and protecting AI agents against prompt-injection attacks during web browsing sessions.

#9 𝕏

Aravind Srinivas announced that Perplexity has open-sourced Bumblebee, a client-side read-only scanner agent for detecting vulnerabilities on macOS and Linux.

#10 𝕏

Philipp Schmid shipped the Interactions API to GA, replacing RPC quirks with standard REST + SSE for a more consistent dev experience; quickstart curl examples and the OpenAPI spec are available in the Gemini API docs.

#11 𝕏

Santiago built a professional voice agent in just 100 lines of Python using LiveKit, Rime, and gpt-4o-mini, delivering one of the most human-like voices and intonations available.

#12 𝕏

Santiago unveiled WorldDiT, a sub-1B-parameter robotics model that both predicts world dynamics and directly generates control actions, proving that much more capability can come in a much smaller package.

#13 𝕏

Hugging Face unveiled Training Agents 3, demonstrating how to train a local, open-weight agent from scratch using reinforcement learning.

#14 𝕏

NVIDIA AI demos Nemotron Labs’ workflow for training open-source models with reinforcement learning on the Prime Intellect platform, covering the end-to-end pipeline and optimization strategies.

#15 𝕏

NVIDIA AI rolled out an end-to-end workflow in the Jetson AI Lab showing how to run GenAI models locally on NVIDIA Jetson devices, leveraging TensorRT and CUDA optimizations for real-time, low-latency edge inference.

#16 𝕏

Rowan Cheung says Mark Zuckerberg runs Meta’s superintelligence lab like a 50–100 person startup, personally recruiting every top AI researcher to keep the whole project in everyone’s head and avoid the outsized drag of underperformers.

#17 𝕏

Cursor launched Cursor Start, a new ₹649/month plan for developers in India that provides generous access to Grok 4.5 and Composer for daily AI agent–powered planning, building, testing, and shipping.

#18 𝕏

Cursor launched Cursor Start with autonomous cloud agents that ship work while you’re away and a Cursor for iOS app to steer them from your phone. It also includes plugins, MCP servers, hooks, and skills to extend your workflow.

#19 𝕏

OpenAI says coding agents now manage everything from routine maintenance to full system redesign, letting scientists dedicate more time to advancing research. Researchers, however, still need to define scientific questions, verify outcomes, and assume long-term ownership.

#20 𝕏

OpenAI warns that accelerating frontier AI development may soon require deliberate pacing to keep it safe and beneficial, and it’s teaming up with the U.S. government, other labs, and the open-source community to build the tools and mechanisms needed to do so.

#21 𝕏

Rowan Cheung shares the full YouTube interview link—watch him unpack AI startup investing, product strategy, and founder playbooks here: https://www.youtube.com/watch?v=WuTJkFvw70o&t=1s

#22 𝕏

Fei-Fei Li unveiled R2S2R’s Real-to-Sim module, which transforms real robots, sensors, environments and interactions into simulations that retain task-relevant visual and dynamic details. This breakthrough sets a new benchmark for sim-to-real alignment in robot manipulation.