Bravr

NVIDIA donated OpenShell, its agent-security runtime, to CNCF: sandboxed tools and policy checks applied as your agents run. When agent security becomes open infrastructure, skipping it is now negligence. blogs.nvidia.com/blog/open-s… #AI #DevOps

n8n just open-sourced a personal agent that signs into websites, runs jobs in parallel and remembers months of context — but every send, pay or delete needs your approval first. Autonomy with the brakes still attached. github.com/n8n-io/nod #AI #Automation

Meta built a model for the boring half of AI agents: tool calls, routing, extraction. Muse Ember costs $0.10 per million tokens, matching the industry floor. The commodity layer is set, so your edge is what you build on top. research.meta.ai/blog/muse-e…

NVIDIA researchers compiled an LLM's reasoning into a CPU-executable cache. Same answers on hard tasks, 16 to 291 times cheaper per call. If your agents keep re-solving the same problems, that is your margin hiding in plain sight. arxiv.org/abs/2611.09891

If your safety scores come from multiple-choice tests, they're flattered. Models commit to an answer letter before refusal ever kicks in, so the benchmark measures knowledge, not safety. Worth knowing before you pick a model on them. arxiv.org/abs/2611.08639

Mistral open-sourced the whole thing: Large 4, all trillion parameters, under a modified MIT licence. It matches closed rivals on coding with 47% fewer tokens per task. Free to download and cheaper to run is a dangerous combination. mistral.ai/news/mistral-larg…

Claude Code's new update moves tool calls into the background and lets you set a rule to auto-deny permission prompts. If you've got agents running, that's the difference between babysitting approvals and actually shipping. aitoolsrecap.com/ai-news-tod… #AI #Agents

Agent-to-agent communication just became production tooling. Google's ADK moved the A2A protocol out of experimental to stable. Building agent systems? You're now on a protocol that won't churn. aitoolsrecap.com/ai-news-tod… #AI

OpenAI is building a ChatGPT Search Console for businesses — first-party impressions, clicks, and citation data for where your brand shows up in AI answers. GEO finally gets its own analytics tab instead of guesses. seroundtable.com/chatgpt-too… #AI #GEO

Your agents can now improve themselves without a single fine-tune. AgentDescent treats prompts and skills as optimisable parameters and runs gradient descent over them, so your system gets smarter as it works. github.com/ThinkFlowLab/agen… #AI #Agents

Hand a coding agent a PyTorch repo and a clean TPU box and ask for a working JAX version. The best model in the world finishes 75% of the time. That is not a tool you hand a production codebase to unsupervised. arxiv.org/abs/2611.08472v1 #AI #CodingAgents

Agents that think in latent space instead of typing out every step cut search time from an hour to minutes. A small 8B swarm even outranks frontier models on some tasks. Your agent bill just got a lot smaller. arxiv.org/abs/2611.08483v1 #AI #Agents

The model you pick isn't the whole story. The harness you wrap around it can swing results by 11.5 points and even flip the model from first to last. Don't judge a model in isolation. arxiv.org/abs/2611.06116 #AI #LLMs

An agent that designs its own serving infrastructure just beat the reference setup by 4.87× on token value per dollar. If your inference bill is a line item, that's not a lab curiosity. arxiv.org/abs/2611.06841 #AI #Inference

Most LLMs answer a question by writing you a paragraph. Sarvam Prism skips that and just returns the probability — 44ms, calibrated, open-weight. If your pipeline only needs a decision, why pay for the essay? sarvam.ai/blogs/sarvam-prism #AI

December 1st: if you're on gemini-2.5-pro or flash, Google auto-routes you to 3.x. thinking_level replaces thinking_budget, sampling params vanish. Check your prompts before your app breaks, not after. ai.google.dev/gemini-api/doc… #GeminiAPI #AI

vLLM's new release goes CUDA 13.0-only, dropping the 12.9 toolchain. If you run your own inference, that's a decision you make before you upgrade, not a minor bump. github.com/vllm-project/vllm… #AI #MLOps

A model that fixes its own weights in plain English, no fine-tuning needed. Researchers edited 130k weights on a small model and MATH-500 jumped from 7.9% to 48.1%. That's the fine-tuning bill gone. alphaxiv.org/abs/2611.weight… #AI #LLMs

Token spend on coding agents has always been a bill you find out about afterwards. Microsoft's TokenEconomizer learns from past sessions and cuts waste by 18-31% on SWE-Bench, without swapping models. github.com/microsoft/TokenEc… #AI #DevTools

Collaborative AI agents are finally learning how to critique and verify their own work before it hits the user. This self-correcting loop is a blueprint for workflows that actually stop the drift. arxiv.org/abs/2610.40589 #AI #AgenticWorkflows