Claude Code Now Pushes Prototypes to Figma Canvas

#1 𝕏

OpenAI launched EVMbench, a benchmark that evaluates AI agents' ability to detect, exploit, and patch high-severity smart contract vulnerabilities on the Ethereum Virtual Machine.

#2 𝕏

Qwen launched the Qwen3.5-397B-A17B-FP8 weights—SGLang support is now merged and a vLLM PR has been submitted—with the model available on Hugging Face and ModelScope and full vLLM integration arriving in the next few days.

#3 𝕏

Google DeepMind rolled out Lyria 3 in beta globally via the Gemini app, enabling easy custom audio generation. Every output now includes SynthID, Google's invisible watermark for AI-generated content.

Also covered by: @Google DeepMind

#4 𝕏

Claude now integrates Claude Code with Figma via the updated MCP server, letting you push working code prototypes directly onto a Figma canvas and explore multiple design versions.

#5 𝕏

Anthropic analyzed millions of interactions across Claude Code and its API to measure how much autonomy users grant AI agents, where they're deployed, and the risks those agents may pose.

#6 𝕏

Cursor now taps into past conversation history as context, enabling more coherent and context-aware AI interactions.

#7 𝕏

Philipp Schmid launched Lyria 3 in the GeminiApp, an advanced AI music model that generates 30-second tracks from text or image prompts with custom lyrics, vocals and cover art.

Also covered by: @Google DeepMind

#8 📝 PromptLayer Blog

Understanding Intermittent Failures in LLMs - The post examines why LLM-based applications that once worked start exhibiting intermittent failures like nonsense outputs, timeouts, or refusals. It emphasizes PromptLayer's observations of production behavior and the challenges of diagnosing non-deterministic, context-dependent issues.

#9 📝 PromptLayer Blog

Prompt Routers and Flow Engineering: Building Modular Self-Correcting Agent Systems - This article describes the move from crafting single prompts to designing full reasoning flows and modular architectures that can self-correct. It explains how teams are shifting from ad-hoc prompt tweaking to systematic designs that catch mistakes automatically.

#10 in

🥞 Carl Vellotti highlights Claude Code's Skills—plain-English, markdown-based automations—and shares 10 rules for building them, including writing trigger-focused descriptions, only automating repeatable tasks, and using slash commands to work around flaky auto-activation.

#11 📝 Eleanor Berger & Isaac Plath

Automating Presentation Slides with Agent Skills - Featured post about agentically creating presentation slides using Slidev, Nano Banana, and Agent Skills. Highlights automating slide creation with agent workflows.

#12 ▶️

This solves agent's context problem - Manage memory like Git

AI Jason

#13 ▶️

How My Claude Code Sonnet 4.6 AI Agent Navigates Chrome Autonomous

All About AI

#14 𝕏

bolt.new rolled out major Bolt for Teams features—team templates, security scans, external database support, admin deploy controls, and model persistence are now live.

#15 𝕏

Alex Tamkin shares new Anthropic-led research (Miles McCain et al.) showing that most AI agents in the wild power software engineering tasks, though adoption is broadening across industries, and overall agent autonomy and deployment risk have both climbed over time.

#16 𝕏

Jason Zhou launched Reusable Components & Pinned Context in SuperDesignDev, enabling project-wide component reuse and pinned design notes for persistent AI context, slashing token use by ~50% and ensuring UI consistency.

#17 𝕏

Teresa Torres shows how AI can synthesize three customer interviews into a draft Opportunity Solution Tree in minutes—then refine it with human expertise, because a usable draft you actually iterate beats a perfect process you never finish.

#18 📝 Simon Willison

Typing - After 25+ years coding, Simon is warming to type hints and strong typing because coding agents can handle the repetitive typing, making explicit types more attractive for tooling and clarity. He notes his prior resistance due to slowed iteration in REPL workflows but now sees potential benefits when agents do the typing work.

#19 𝕏

NVIDIA AI: AI Dungeon by Latitude now runs large-scale mixture-of-experts models on the DeepInfra inference platform with NVIDIA Blackwell GPUs (NVFP4, TensorRT LLM) at just $0.05 per million tokens (down from $0.20).

#20 𝕏

LlamaIndex 🦙 is running a LlamaAgents contest and released a new walkthrough for its LlamaAgent Builder. By describing a document workflow in natural language, it auto-selects and configures LlamaSplit + LlamaExtract to generate a deployable agent with API and UI.

#21 𝕏

NVIDIA AI leverages edge computing with RideAct and Hayden AI to automate bus-lane enforcement, slashing transit delays and boosting accessibility for millions of riders—all while upholding strict privacy standards.

#22 𝕏

Peter Yang argues that as AI agents take the first pass on everything, products, platforms, and even internal docs must shift from user-centric UIs to being fully agent-optimized.

#23 𝕏

Dharmesh Shah relays Vercel founder Guillermo Rauch's advice to be intentional about building AI projects. He frames AI as an intellect amplifier and mirror of your values, warning against confirmation bias and overbuilding in the AI era.

#24 in

Anu Jagga Narang proposes using "Can you do it in the same day?" as a progress metric for AI and warns that as automation accelerates, the real edge lies in choosing the right projects and building agile organizational systems.

#25 in

Scott Brinker turned feedback on his context-as-a-service newsletter into a 3,000-word deep dive—sharing Zylo tech-stack data, AI-agent use cases, UI-decoupling nuances, monetization models, and theory rebuttals.