Deep Currents 06.06.26

Reading the Currents

Agents Everywhere

This month the agents began their invasion of the desktop. Microsoft introduced Scout, an always-on agent that manages email and meetings and flags stalled decisions on its own, while Google's Gemini Spark and Nous Research's desktop version of Hermes launched with the same 24/7 premise. And while those platforms rely on cloud services to run, some agents are moving closer to home. Perplexity released a Mac agent under the banner "the data centre moves to your machine" and new models keep shrinking to fit on smaller computers: Google's Gemma 4 runs on a 16GB laptop, Cohere's Command A+ on two GPUs, and a bunch of open-weight releases (MiniMax's M3, NVIDIA's Cosmos 3 and Nemotron 3 Ultra) put real capability in the hands of anyone with the hardware to run it.

Show Me the Money

Another big change is we're being asked to hand over our finances to AI. Robinhood launched agentic trading and an agentic credit card, letting an agent place trades and carry a balance for you. Meanwhile Anthropic shipped ten financial-services agents, Perplexity built 35 finance workflows, and OpenAI added personal finance to ChatGPT. The capabilities were mostly there already, but now we're seeing products and connectors coming out that are tailored specifically to handle our money.

Two Tiers

The capabilities are also splitting in two as they spread, as I've previously predicted. The everyday stuff is becoming more accessible, able to run on laptops and available to download with open weights. The most advanced stuff is going the other way. OpenAI released GPT-5.5-Cyber behind a restricted trusted-access program, modelled on Anthropic's similarly restricted Mythos model, on the logic that a model that's good enough to penetrate critical software's defences should not circulate freely.

Going Public

And while all this is going on, SpaceX, Anthropic, and OpenAI are all getting ready to go public with blockbuster IPOs over the next months, with some economists predicting three pricks that pop the AI bubble.

Okay, on to the comprehensive list of news and announcements from the past month...

The Full Stream

Agents

• Anthropic announced a handful of new features to improve complex work handled by the Managed Agents platform for Claude Code, including Dreaming to extend memory by reviewing past sessions to find patterns and help agents self-improve, and self-hosted sandboxes and MCP tunnels. Claude Managed Agents also got self-hosted sandboxes and MCP tunnels for safer enterprise API connections. Meanwhile Claude Code got dynamic workflows, which can split big coding jobs across multiple subagents to burn through your token budget even faster.
• Anthropic unveiled 10 ready-to-run Claude agents for financial services and insurance, capable of handling work ranging from building pitchbooks and screening KYC files to reviewing earnings and valuations, and launched Claude for Small Business with a handful of new connectors for tools that small businesses frequently use.
• OpenAI released a Codex for Chrome plugin that lets Codex autonomously drive background tabs on Mac and Windows for deep research, CRM data transfer, and admin-console workflows.
• Cursor shipped a /orchestrate skill that lets a parent agent spawn and coordinate child agents recursively for big refactors. They also launched cloud development environments that make it easier for teams to run multiple parallelized agents to handle tasks end-to-end, inside separate cloud-hosted development environments.
• Cursor released Composer 2.5, improving long-running coding-agent behavior, instruction following, and cost efficiency versus Composer 2.
• Google's Antigravity 2.0 orchestrates parallel coding agents, background tasks, CLI workflows, and SDK access.
• Google launched Gemini Spark, a 24/7 personal agent that can work across Workspace apps.
• Higgsfield launched Supercomputer, an agentic creative pipeline that automates social media posting, monitors mentions and replies, and deletes anything off-brand.
• Kimi WebBridge connects Kimi's desktop agent to your browser so it can click, fill forms, and complete web tasks.
• Microsoft released Copilot Cowork, an agent that can take a complex goal, build a plan, schedule meetings, draft documents, and execute across your entire Microsoft 365 account. They also released iOS and Android apps.
• Microsoft launched Scout, an always-on agent powered by OpenClaw and available through Microsoft's Frontier program.
• Nous Research released a Hermes desktop app for their always-on agent platform.
• OpenAI gave Codex several new role-specific plugins to address common business needs, including data analytics, product design, and sales, with more coming soon. They also launched Sites, allowing users to share their Codex creations as hosted websites or apps, and added an annotation feature that lets you select a specific component and type your instructions to change.
• Perplexity released Personal Computer for Mac, a desktop app that browses, takes actions, and runs research workflows on your behalf locally, and launched Computer for Professional Finance, bringing licensed data and 35 dedicated workflows to its agentic system to help analysts handle routine work.
• Perplexity's Computer can now split agent tasks between a local model on your device and frontier models in the cloud.
• Robinhood launched Agentic Trading and an Agentic Credit Card, letting users connect AI agents to execute stock trades, manage spending, and automate purchases.
• xAI launched Grok Build, its rival to Codex and Claude Code, in beta.

Design Tools

• Figma Make now connects to your production codebase, allowing you to make edits to your Make prototype and issue pull requests for your engineering team to review.
• Paper launched SVG generation and PDF exports (requires Paper Desktop), and now fully supports copy/paste from Figma, including images and SVGs. They also added vector editing and a pen tool.
• Pencil added a new "Let it cook" feature that automatically iterates on your designs — choose between Layout and Style and it will explore up to 6 variations. They also added "quick actions" you can apply to any top-level frame or multiple frames, including layout variation, style variation, explore typography, and more.
• Replit added a Canvas mode that lets you iterate on layouts, and draw and annotate your designs to specify changes, then apply those changes to your app.

Frontier Models

• Alibaba previewed Qwen 3.7 Max, an agent model built for long work sessions, and then released Qwen3.7-Plus, a multimodal agent model that unifies vision and language.
• Anthropic released Opus 4.8, putting it back on top of the benchmarks at the same price as 4.7. They also added effort controls (on top of adaptive thinking) so you can choose the level of thinking you need for any given task. Anthropic also signed a compute partnership with xAI to address their struggles meeting demand.
• Cohere launched Command A+, a new open-source agentic model designed for enterprise clients requiring data sovereignty, able to run on two H100 GPUs.
• Google released Gemma 4 12B, a new multimodal model able to run on a 16GB laptop, and the first Gemma variant of this size built for native audio.
• Microsoft released seven new models under a new MAI label.
• MiniMax released M3, calling it the first open-weight model to pair top-tier coding with a 1M-token context window and native image and video input.
• NVIDIA released Cosmos 3, an open-weight foundation model built for robots and self-driving. It combines reasoning, world generation, and action prediction into a single system.
• NVIDIA also released Nemotron 3 Ultra, a 550B parameter open weights model that competes with Qwen3.5 and Kimi K2.6.
• OpenAI announced GPT-5.5-Cyber, a cybersecurity-focused model with similarly restricted access to Anthropic's Mythos.
• OpenAI's GPT-5.5-Instant started rolling out to all ChatGPT users, bringing improved performance, stronger memory, and more personalized, concise responses.
• Thinking Machine Labs released a new model that emphasizes fluid interaction .
• Trajectory, a new startup founded by ex-DeepMind and Apple researchers, launched with $15M to build the platform for continual learning so the AI gets smarter from real-world experience.

Images

• Google launched Google Pics, which edits image objects, text, and Workspace visuals with precise controls.
• Ideogram launched a new background remover then rebranded with a new logo and released version 4.0 with open weights, meaning you can download the weights, fine-tune on your own data, and run it on your own hardware.
• Krea released Krea 2, a new image generation model built from scratch, focused on aesthetics.
• Leonardo released a 3D model tool to turn 2D images into 3D characters.
• OpenAI is adopting SynthID (Google's watermarking system) for ChatGPT images, alongside launching a public verification tool.
• Recraft launched V4.1 calling it their most beautiful model ever. 
• Reve released Reve 2.0, claiming to be the best 4K image model in the world. The new model uses object segments and layers to provide precise editing control.

Music

• ElevenLabs released Music v2, upgrading its music-generation model with better vocals, instrumentation, multilingual support, mid-track genre shifts, and track-level inpainting.

Policy & Regulation

• California Governor Gavin Newsom signed an executive order directing state agencies to study and develop policies around how to protect workers from AI-driven job losses.
• Canada released a new national AI Strategy called AI for All that will invest billions of dollars into programs and infrastructure. 
• The European Commission released draft high-risk AI Act guidance and opened consultation through June 23.
• OpenAI published its Frontier Governance Framework, explaining how its safety and security practices map to emerging AI laws.

Productivity

• Google AI Studio now builds native Android apps from prompts and lets you connect to Gmail, Drive, and Sheets.
• Google's NotebookLM rolling out automatic Google Drive file sync, so your notebooks stay updated when source docs change.
• Notion announced a developer platform that enables you (and your AI coding agents) to sync any data source, build custom agent tools, and trigger Notion workflows.
• OpenAI released a personal finance feature for ChatGPT (U.S. only).

Video

• Pika Labs released a new MCP connector and a handful of useful skills starting with Podcasts, Explainer videos, and UGC ads that you can run in Claude, Codex, etc.
• Runway released a new MCP connector for generating videos from Claude and ChatGPT, and launched Aleph 2.0inside the new Edit Studio — edit a single frame in your video, preview the change, and have it propagated across the rest of your video.

Voice

• ElevenLabs released Dubbing v2, a model that translates video across 90+ languages while preserving the original speaker's tone, performance, and delivery.
• ElevenLabs added Templates for workflows to their Creative suite.
• Miso Labs launched Miso-TTSan open source text-to-speech model that responds in just 110ms, faster than the latency of human conversations. 

World Models

• Odyssey launched two new world models: Starchild-1 for real-time audio/video world simulation, and Agora-1 for shared multiplayer worlds.

Okay, that's enough for another month! I'm not sure if I'm going to keep doing these, so please give this post a thumbs up or reach out and let me know if you are getting value from these monthly recaps, or if you have questions or thoughts to share, or need any help making sense of all this.

Cover image created with Midjourney 8.1. Editing assistance provided by Claude Sonnet 4.6 and Opus 4.8.