Deep Currents 02.02.26

This month brought a tidal wave of agent-related announcements, a viral crustacean, and the clearest signal yet that the big labs are competing not just on model capability but on goodwill and trust. If January was about consolidation, February is about execution: companies are shipping the agentic future they've been promising.

If you want the full rundown of updates, scroll down to the complete listing. Otherwise, here's what I think matters most from this month's developments...

Reading the Currents

The Claw That Broke the Internet

An AI-powered lobster went viral last week. OpenClaw, an open-source self-hosted agent that can control your browser, execute terminal commands, send you messages, and build itself new capabilities on the fly, captured the internet's imagination with its crustacean mascot and impressive, if somewhat alarming, feature list. Until a few days ago it was (briefly) called Moltbot, after changing its name from Clawdbot, which Anthropic's legal team felt sounded too much like Claude-bot. The hasty double rebrand tells you something about the velocity of this moment.

We're watching agents graduate from concept to product. The pattern was proven in the coding tools first: Claude Code, Cursor, Windsurf... Developers discovered that giving an AI the ability to actually do things (write files, run commands, iterate on its own output) unlocked capabilities that chat alone never could. Now that pattern is spreading fast. Anthropic launched Claude Cowork, a version of Claude Code designed for non-developers, inspired by the growing number of people who began using the coding tool to automate non-coding tasks. Airtable launched Superagent. Minimax released their Agent platform. Google shipped Agentic Vision, which gives Gemini a "think, act, observe" loop for analyzing images. Openwork arrived as an open source option, and then OpenClaw dropped...

What's particularly interesting for those of us in the design world is how quickly these agent capabilities are merging with design workflows. Pencil launched as an "agent-driven MCP canvas" that works directly inside IDEs like Cursor and Claude Code. Southleft published documentation for their Figma Console MCP, which lets AI assistants access your Figma design system like an API. Magic Patterns added OAuth integration to their MCP server. The walls between vibe coding and vibe designing are coming down. For designers who've been watching developers talk about agentic workflows from the sidelines, the tools to join that conversation are finally arriving.

Racing to Own the Transaction

Meanwhile, ecommerce is racing to become agentic too. OpenAI integrated shopping capabilities into ChatGPT months ago, and now the rest of the industry is scrambling to catch up. Google launched the Universal Commerce Protocol (UCP), an open standard for agentic commerce that plays nicely with other emerging protocols like Agent2Agent and the Agent Payments Protocol. Microsoft shipped Copilot Checkout, letting users make purchases directly within chat, with Shopify, Stripe, and PayPal integrations at launch. They also released Brand Agents, giving retailers AI-powered customer guidance that adopts their brand voice. The infrastructure for AI-mediated transactions is being built out rapidly.

The Societal Pitch

Both Anthropic and OpenAI made significant healthcare plays this month. Claude for Healthcare can speed up prior authorization reviews, support claims appeals, coordinate care, and connect personal health data through integrations with Apple Health, Android Health Connect, and other services. ChatGPT Health offers a separate, more private channel for health questions, developed with input from 260 physicians across 60 countries. OpenAI also announced HIPAA-compliant offerings for healthcare institutions.

Education saw similar attention. Anthropic partnered with Teach For All to bring AI tools and training to educators in 63 countries. Google Classroom added free practice SATs, new teaching dashboards, and a podcast-style audio lesson format. OpenAI launched Education for Countries as part of their initiative to work with governments and universities on AI in education. Stanford opened AI4ALL to teach ninth graders about AI through hands-on research projects.

Reading these announcements generously, the big labs are investing in sectors that could genuinely benefit society. Reading them cynically, healthcare and education are what you point to when someone asks whether all those data centers and ad-supported chatbots are actually good for humanity. Both readings are probably true. What matters is whether the tools actually help teachers and patients, not whether they help the labs' reputations.

Values in Contrast

The month also delivered a sharp contrast in how companies are approaching the responsibility question. Grok launched unrestricted image generation capabilities, and international backlash over non-consensual deepfakes followed almost immediately. The scandal prompted US politicians to demand Apple and Google remove both X and Grok from their app stores. X.ai restricted the feature to paying subscribers, but the damage was done.

Anthropic, meanwhile, published Claude's Constitution: a lengthy document defining the values and behaviors the company hopes Claude will always follow. Written for Claude as its primary audience and released under a Creative Commons license, it's an unusual artifact of transparency in a field not known for it. The UK government's selection of Anthropic to help build an AI assistant for GOV.UK feels like a real-world signal that the approach is resonating where trust matters most.

The Full Stream

Okay, now for the full rundown of releases across every category that mattered this month...

Agents

Airtable launched Superagent, a coordinated multi-agent AI platform focused on the business market, able to plan and conduct research, perform analysis, and generate sophisticated reports.

Anthropic launched Claude Cowork, a sequel to Claude Code designed for non-developers. It was built in just 1.5 weeks using Claude Code, initially rolling out as a research preview for Max subscribers before expanding to all paid plans. Not long after they added support for plugins, which lets users bundle any skills, connectors, slash commands, or sub-agents together to perform complex work tasks. Anthropic also extended Claude in Excel to Pro subscribers, meaning all paid accounts can now use this agent plugin to take action on spreadsheets, and they made Claude Code available to standard Team license users, so now all paid accounts have access.

Google released Agentic Vision for Gemini 3 Flash, which uses a multi-step "think, act, observe" approach when analyzing images. The model plans its analysis, uses tools to zoom, crop, and annotate, then makes a final assessment, resulting in 5-10% accuracy improvements.

Minimax launched Minimax Agent, a platform for building custom agents from prompts or choosing from predefined Experts like Tidy Folder and Icon Builder. A desktop app version is also available.

OpenClaw (briefly branded Moltbot, but originally known as Clawdbot) is an open-source self-hosted agent you can run locally and talk to from any messaging app including WhatsApp, Telegram, iMessage, Slack, and Discord. It can control your browser, execute terminal commands, install skills, send you messages, remember everything, act autonomously, and build itself new capabilities on the fly. Use with caution. Seriously, it's a potential security nightmare.

OpenWork is an interesting alternative to Claude Cowork. It's an open-source AI agent that runs on Mac, connects to any LLM via API key, and only accesses folders you explicitly grant permission to (where it can read files, create documents, and otherwise automate repetitive tasks.)

Scouts deploys AI agents that continuously monitor the web 24/7 for recurring tasks like house hunting, job searches, or tracking niche news.

Vercel launched an open directory of publicly available Skills for agents, organized like a leaderboard showing installation counts for the most popular and trending.

Audio & Music

Eleven Labs released an album of songs composed by artists using its music generation model. Artists including Liza Minnelli and Art Garfunkel are featured across genres spanning rap, pop, R&B, EDM, cinematic scoring, and "global sounds." Some tracks are entirely AI-generated while others blend AI instrumentals with cloned voices from ElevenLabs' licensing marketplace. It's available as a playlist on Spotify.

Spotify is rolling out Prompted Playlists to the US and Canada after beta testing in New Zealand since late last year. Basically it's an AI chatbot that can generate playlists based on your request, similar to ChatGPT.

Design

Magic Patterns, a platform that helps designers create mockups with real code from a prompt, by entering a URL, or by importing a Figma file, launched OAuth 2.0 integration for their MCP server, making it easier to integrate with Cursor or Claude Code.

Pencil is a new design app described as an agent-driven MCP canvas. It works directly in IDEs like Cursor, VS Code, Codex, and integrates with Claude Code, leveraging curated component kits, and supporting copy-pasting from Figma. Fun!

Southleft launched documentation for their Figma Console MCP tool, which lets designers and engineers collaborate by using AI assistants to access Figma design system files like an API.

Ecommerce

Google launched the Universal Commerce Protocol (UCP), a new open standard for agentic commerce built to work across verticals and compatible with Agent2Agent (A2A), Agent Payments Protocol (AP2), and Model Context Protocol (MCP).

Microsoft launched Copilot Checkout, allowing users to make purchases directly within chat. Payment integrations include Shopify, Stripe, and PayPal, with Etsy shops, Urban Outfitters, and other major retailers supported at launch. Microsoft also released Brand Agents for retailers, offering AI-powered customer guidance that adopts brand voice directly within Shopify sites.

Education

Anthropic announced a partnership with nonprofit Teach For All to bring AI tools and training to educators in 63 countries. The program includes a collaborative component where Anthropic solicits input directly from educators to improve tools for teachers.

Google Classroom now offers free practice SATs, new dashboards to boost teaching and learning, and a podcast-style audio lesson format for teachers to complement existing materials.

OpenAI launched Education for Countries as part of their OpenAI for Countries initiative, working with governments and universities to personalize learning, reduce administrative burden, and prepare students for the workforce.

Stanford launched AI4ALL to teach 9th graders about AI through hands-on research projects and mentorship from Stanford researchers, available online and through residential summer programs.

Government

Anthropic was selected by the UK's Department for Science, Innovation and Technology (DSIT) to help build and pilot an AI-powered assistant for GOV.UK. The chatbot will help people navigate government services with tailored advice, starting with employment assistance to help people find work, access training, and understand available support.

Health

Anthropic announced Claude for Healthcare, which can speed up prior authorization reviews, support claims appeals, coordinate care and triage patient messages, support healthcare startups, and connect personal health data through integrations with HealthEx, Function, Apple Health, and Android Health Connect. Four health integrations are now in beta in the Claude iOS and Android apps for Pro and Max subscribers.

OpenAI launched ChatGPT Health as a separate, more private channel for health questions, medical record uploads, and health API connections including Apple Health. It was developed with input from 260 physicians across 60 countries and numerous specialties, covering 600,000 outputs across 30 focus areas. OpenAI also launched ChatGPT for Healthcare with HIPAA compliance for US healthcare institutions.

Images

Black Forest Labs released Flux.2[klein], their fastest image generation and editing model to date, available in two sizes with the smaller 4B variant able to run on consumer hardware with as little as 32GB of RAM.

KREA AI introduced Realtime Edit, which lets you edit images with complex instructions in real time.

ImagineArt released ImagineArt 1.5 Pro with native 4K output.

Midjourney released Niji 7, an updated model specifically designed for anime-style images with better support for Sref codes than their main model. They also updated the Style Creator tool with improvements including the ability to start from an existing Sref code or text prompt.

Runway released Story Panels, an app that turns a single image into a three-panel storyboard.

Bytedance released Seedream 4.5 which can create cinematic images from text with accurate spatial layout and world knowledge, up to 4K resolution.

X.ai launched and then quickly restricted Grok's image generation to paying subscribers only after international backlash over non-consensual deepfake images. The scandal prompted US politicians to demand Apple and Google remove both X and Grok from their app stores.

Z AI released GLM-Image, an open-source image generator hailed as the first major model trained entirely on Huawei hardware.

LLMs

Alibaba released Qwen3-Max-Thinking, a new flagship reasoning model competing with ChatGPT-5.2, Gemini 3 Pro, and Claude Opus-4.5.

Amazon launched Alexa+ as a web-based chatbot to challenge ChatGPT and Gemini. alexa.amazon.com/about

Anthropic published Claude's Constitution, a document defining the values and behaviors Anthropic hopes Claude will always follow. It's lengthy, written for Claude as its primary audience, and released under a Creative Commons license. Anthropic also added support for interactive tools in Claude based on the new MCP Apps protocol.

Apple reportedly signed an agreement with Google to use Gemini for its next-generation Siri.

Arcee released Trinity Large Preview as an open-source non-reasoning model comparable to Llama 4 Maverick on benchmarks, with a full reasoning version coming after further post-training.

Google launched a $7.99/month AI Plus plan in the US and 35+ countries including Gemini 3 Pro, NotebookLM, Flow's AI filmmaking tools, and 200GB storage. Unlike OpenAI's competing Go plan, it won't have ads. Google also announced personal intelligence, giving Gemini access to all connected Google services from Gmail to YouTube for Pro and Ultra subscribers in the US.

Kimi released K2.5, featuring an Agent Swarm capability that spawns up to 100 parallel sub-agents across 1,500+ tool calls to tackle complex tasks.

Liquid AI released LFM2.5, an open-weight model family designed to run AI directly on devices without cloud reliance. The lineup includes text, vision, audio, and Japanese variants, with the audio model 8x faster than its predecessor and the 1.2B text model outperforming Llama 3.2 and Gemma 3 on instruction following, tool use, and math.

NotebookLM can now create Data Tables from your notes and sources, structured and exportable to Google Sheets.

OpenAI expanded ChatGPT Go worldwide at $8/month with most features but lower limits and no GPT-5.2 Pro access. They also announced an advertising model for free and Go plan users, with ads appearing below responses for US users over 18. The final format may include the ability to "chat" with advertisers. Finally, they announced they’ll be retiring GPT-4o for good on February 13. Last summer when they released GPT-5 they tried to retire it but faced a huge backlash, so we’ll see what kind of an outcry there is this time.

TII Abu Dhabi released Falcon H1R, a 7B-parameter model matching larger reasoning systems on math, coding, and general benchmarks using a hybrid Transformer-Mamba2 backbone with 256k context window.

Z AI released GLM-4.7-Flash, a 30B parameter variant balancing performance with efficiency.

Research & Safety

AVERI (AI Verification and Evaluation Research Institute) launched with $7.5M in funding to push for independent safety audits of frontier AI models.

Anthropic released an updated Economic Index report examining changes in AI use globally and across the US since September 2025.

The UK's AI Safety Institute published a frontier AI trends report based on two years of analysis.

Science

OpenAI released Prism, a free AI-native workspace for scientists to write and collaborate on research.

Search

Yahoo! is back. They launched Yahoo Scout, an AI-first "answer engine" leveraging Yahoo's proprietary data and 30+ years of search history, partnering with Anthropic to use Claude as Scout's primary foundational model.

Text & Translation

DeepSeek open-sourced OCR 2, a document text extraction model that tops benchmarks while being more token-efficient.

Google released TranslateGemma, a new collection of open translation models built on Gemma 3.

Locally Translate is an iOS app from Locally AI that runs completely offline using the iPhone's on-device processors and MLX framework for private translations in 50+ languages.

OpenAI launched ChatGPT Translate, a dedicated translation tool.

Vibe Coding

AI2 introduced SERA, a new family of open-source coding agents designed to be trained on private codebases, with setup scripts and inference optimizations compatible with Claude Code.

Anthropic released Claude Code 2.1 with a slew of updates and improvements, and also made it accessible via the Claude Desktop app, allowing users to easily toggle between regular Chat and Code. Key additions include hooks for agents and skills, hot reload for skills, and wildcard tool permissions. They also updated Claude Code with MCP Tool Search, enabling “lazy loading” of external tools, thereby reducing context usage from 134K to 5K tokens. 

Cursor rolled out dynamic context discovery for more token-efficient context management. Version 2.4 adds subagents for codebase research and parallel work streams, Agent Skills support, image generation with Nano Banana Pro, and PDF reading in chats. They also added secure codebase index sharing to reduce setup time on large codebases.

Lovable launched 4 new features that add up to a big release: Plan mode, prompt queuing, automated testing with browser tools, and Google sign-in integration for apps.

Mistral launched Mistral Vibe, a coding tool similar to Claude Code running on Codestral and Devstral models. Available in terminal, as an IDE extension, or self-hosted.

Moonshot released an open-source version of Kimi Code, their agentic coding agent for terminals and IDEs.

Replit launched mobile app creation, allowing users to build, test via QR code, and publish directly to the App Store.

Windsurf released Wave 13 with parallel multi-agent workflows, Git worktrees, and a dedicated terminal. Their SWE-1.5 coding model is free for everyone for the next three months.

Video

Adobe rolled out new AI features for Premiere including one-click masking with Object Mask and Firefly Boards integration for seamless storyboard-to-video workflows.

Google added Ingredients to Video with Veo 3.1, combining multiple reference images with 4K upscaling. Google Labs also opened Flow to all Workspace users, though credits are limited at most subscription levels.

LTX open-sourced LTX-2, an AI video model generating native 4K footage with synced audio and granular camera/motion control. They also released an audio-to-video tool with Eleven Labs for synchronized video from still images and audio files.

Luma Labs released Ray3.14, their most professional model with native 1080p generation, 4x faster and 3x cheaper.

Remotion launched a Skill for Claude Code to generate videos using their open-source React framework from text prompts.

Wondercraft pivoted from audio to video, and launched Wondercraft Studio, along with an agent called Wonda, that creates full videos (not just clips) starting from a text prompt.

xAI released the Grok Imagine API that offers text-to-video, image-to-video, and image editing with low latency and at a low cost.

Voice & Transcription

Eleven Labs released Scribe v2, their next-generation transcription tool with cleaner captions and faster processing, available via API or ElevenLabs Studio.

Every launched Monologue, a multilingual voice-to-text tool similar to Wispr Flow with auto-formatting, auto-dictionary, and auto-editing. Runs locally on M-series Macs or in the cloud.

FlashLabs open-sourced FlashLabs Chroma 1.0, a near real-time speech model with personalized voice cloning.

Nvidia released Nemotron Speech ASR, an open-source transcription model designed for low-latency voice agent use cases.

Willow is another Wispr Flow alternative for dictation anywhere (including offline) that auto-learns names and terms.

Web Browsers

Google updated Chrome with a Gemini sidebar featuring Autobrowse and UCP ecommerce protocol integration for AI-assisted shopping. They also added image generation with Nano Banana Pro for editing images on the web.

Norton launched Neo, a safety-first AI browser that stores chats and browsing history locally, lets users control what Neo remembers, and includes Norton Web Shield for blocking malicious sites and phishing attacks.

World Models

Google officially launched Project Genie, a web app they first previewed last summer, that lets users create and explore 3D worlds. Available exclusively to Google AI Ultra subscribers in the U.S. for now.

Odyssey launched Odyssey-2 Pro, a world model streaming real-time interactive video at 720p/22fps. The Pro model is significantly larger, faster, and more realistic than the original, now available through an API.

PixVerse released R1, a real-time world model that understands physics and object permanence for consistent, realistic motion.

World Labs launched the World API for generating explorable 3D worlds from text, images, and video.

If you made it this far, congratulations on your dedication to keeping up with the relentless pace of generative AI! As always, please reach out if you have thoughts to share or stories I should be tracking.

Cover image created with Midjourney.