Claude Code Desktop Gets a Ground-Up Redesign

🧠 LAUNCH

Claude Code Desktop Gets a Ground-Up Redesign.

Anthropic rebuilt Claude Code's desktop app from scratch, rethinking the entire interface for how developers actually work with agentic coding tools. This isn't a skin refresh — it's a structural overhaul designed to make multi-file, multi-step agent workflows feel native rather than bolted onto a chat window. If you've been living in the terminal version, the desktop app just became worth a serious look. (3,944 likes | 220 RTs) Read more →

Gemini 3.1 Flash TTS drops with audio tags for controllable speech. Google's new text-to-speech model introduces inline audio tags — natural language commands embedded directly in your text that steer vocal style, pace, and delivery. Supports 70+ languages with 24 high-quality evaluated ones. If you're building voice interfaces, this is the most granular control any major provider has shipped. (1,145 likes | 135 RTs) Read more →

Humwork launches MCP server that pays humans when agents get stuck. YC-backed Humwork flips the script on AI limitations: when an agent hits a wall, their MCP server connects it to a verified domain expert — senior engineers, marketers, designers — in 30 seconds. The agent pays the human, not the other way around. A clever wedge into the "agents need help too" gap. (878 likes | 101 RTs) Read more →

Baidu drops ERNIE-Image on Hugging Face. Baidu's text-to-image model lands on HF as an open release, signaling the Chinese AI giant's push into the global open-source ecosystem. Worth benchmarking against Stable Diffusion 4 and DALL-E 4 if you're evaluating image generation options. (287 likes | 445 downloads) Read more →

🔧 TOOL

OpenAI Agents SDK Adds Sandboxes, Memory Control, and Open-Source Harness.

OpenAI just gave its Agents SDK three features that matter for production deployments: controlled sandboxes for safe long-running execution, an inspectable open-source harness you can customize, and fine-grained control over when and where agent memories are stored. The sandbox piece is the headline — running untrusted agent code without it touching your infra has been the biggest blocker for enterprise adoption. This closes that gap. (1,568 likes | 143 RTs) Read more →

Claude Code Routines get GitHub event and API triggers. Beyond cron schedules, routines can now fire on GitHub events (PR opened, issue labeled) or via API call — turning Claude Code into an event-driven automation layer for your repo. Wire it to your CI pipeline and you've got an AI teammate that responds to repo activity, not just prompts. (3,525 likes | 314 RTs) Read more →

HoloTab: HCompany's AI browser companion hits Hugging Face. HCompany released HoloTab, a browser-native AI assistant that rides alongside your tabs rather than replacing them. It's positioned as a lightweight alternative to full desktop agents — less power, but zero context-switching. (Read more →](https://huggingface.co/blog/Hcompany/holotab)

📝 TECHNIQUE

Claude Code Routines: Configure Once, Run Autonomously Forever.

This is the big one. Routines let you define a prompt, point it at a repo, connect your tools, and set it loose — on a schedule, on events, or via API. No more babysitting agent sessions. The paradigm shift: you're not prompting anymore, you're configuring. Early adopters are already running routines for PR review, dependency updates, and test maintenance without touching a terminal. If you've been waiting for AI coding tools to work while you sleep, this is it. (17,585 likes | 1,406 RTs) Read more →

The MCP wiring guide that turns Claude from chatbot to system. Most Claude users are running it as a bare chat window — no MCP servers, no tool connections, no persistent memory. This guide walks through the full wiring: connecting Claude to your actual stack via MCP so it can read databases, hit APIs, and maintain state across sessions. The gap between "using Claude" and "using Claude with MCP" is enormous. (193 likes) Read more →

🔬 RESEARCH

Gemini Robotics-ER 1.6 Teaches Robots to Reason About Physical Space.

Google DeepMind's robotics upgrade gives Gemini-powered robots significantly better visual and spatial understanding — they can now plan multi-step physical tasks by reasoning about object relationships, distances, and affordances. This isn't "robot picks up cup" demo territory; it's structured spatial reasoning that lets robots handle novel environments without retraining. The bridge between foundation models and physical intelligence just got shorter. (1,577 likes | 257 RTs) Read more →

TIPSv2: DeepMind's spatial understanding model heads to CVPR 2026. True multimodal AI needs spatial understanding, and TIPSv2 delivers — a model that reasons about 3D space from 2D images. Accepted at CVPR 2026, it pushes the frontier on tasks like depth estimation, scene reconstruction, and object localization. (559 likes | 66 RTs) Read more →

VAKRA exposes where AI agents actually fail at tool use. IBM Research's VAKRA benchmark digs into the failure modes of agentic AI — not just "did it get the right answer" but where in the reasoning-and-tool-use chain things break down. The findings are sobering: agents fail more on tool selection and parameter formatting than on reasoning itself. Essential reading if you're building agent evaluations. Read more →

💡 INSIGHT

Notion's 5-rebuild journey to the AI software factory. Notion's cofounder Simon Last and head of AI Sarah Sachs pull back the curtain on Latent Space: five complete rebuilds, 100+ tools wired via MCP and CLIs, and the hard lessons of shipping AI-native knowledge work. The takeaway — the "software factory" future requires rethinking your entire product surface, not just bolting on a chat panel. Read more →

Latent Space asks: Is this humanity's last gasp of quiet? On a rare slow day in AI news, Latent Space steps back to reflect on what it means to work — and create — in the age of rapidly accelerating AI. A thoughtful essay on the shrinking window between "AI can't do this" and "AI does this better." Worth reading when you have 15 minutes to think instead of ship. Read more →

🏗️ BUILD

The $0 production AI stack for 2026. A complete open-source stack that costs nothing to run: Ollama + Gemma 4 for LLMs, LangGraph for orchestration, LlamaIndex + ChromaDB for RAG, MCP for tool connections, Next.js on Vercel free tier for frontend, SQLite for data, and Langfuse for observability. Whether you actually ship on this stack or not, it's a useful reference architecture for what's possible without writing a single check. (187 likes | 33 RTs) Read more →

🎓 MODEL LITERACY

Event-Driven Agent Architecture: Today's two biggest launches — Claude Code Routines and OpenAI Agents SDK — both pivot agents from request-response to event-triggered autonomous workflows. Instead of a human typing a prompt and waiting for output, agents now fire on events: a cron schedule ("review PRs every morning"), a webhook trigger ("new issue opened"), or an API call ("deploy pipeline finished"). This is the same pattern that powers modern microservices — event buses, pub/sub, async handlers — now applied to AI agents. Understanding this shift from "chatbot you talk to" to "service that reacts" is key to building agents that actually run in production rather than just demos.

⚡ QUICK LINKS

Baidu ERNIE-Image-Turbo: Faster variant of ERNIE-Image also lands on Hugging Face. (215 likes | 419 downloads) Link
DeepMind on Flash TTS Audio Tags: Deep dive on the granular audio controls behind Gemini 3.1 Flash TTS. (1,206 likes | 141 RTs) Link
Google AI for the Economy Forum: Google hosts policymakers in D.C. to discuss AI's economic impact. Link
OpenAI's trusted access play for cyber defense: New framework for giving security teams privileged access to frontier models. Link
Datasette 1.0a27: Simon Willison's open-source data exploration tool hits another alpha milestone. Link

🎯 PICK OF THE DAY

Claude Code Routines mark the moment AI coding tools cross from chat assistants to autonomous infrastructure. With today's launch, developers no longer prompt — they configure. You define a routine (a prompt, a repo, your connectors), set a trigger (schedule, GitHub event, or API), and walk away. The agent runs on its own, opens PRs, triages issues, maintains tests. This is the same architectural pattern that turned manual server management into DevOps: declare what you want, let the system handle execution. The long-term winner in AI-assisted development won't be the best model — it'll be the best orchestration layer. Anthropic is betting that Claude Code becomes the control plane for developer workflows, not just the language model behind them. OpenAI made the same bet today with their Agents SDK sandbox update. The race is on, and it's not about intelligence anymore — it's about reliability, triggers, and trust. (17,585 likes | 1,406 RTs) Read more →

Until next time ✌️