🤖 Claude now runs 1,000 coding agents at once

Good Morning! Here is today’s breakdown:

Anthropic released Claude Opus 4.8, just 41 days after Opus 4.7, holding pricing flat at $5/$25 per million tokens while adding Dynamic Workflows, a new Claude Code feature in research preview that lets the model orchestrate up to 1,000 parallel subagents in a single session. It is available now on claude.ai, the API, Amazon Bedrock, and Google Vertex.

Dynamic Workflows changes what agentic coding actually means. Before this, an agent handled one file at a time while a human stitched the sessions together. The 750,000-line Bun migration in 11 days is not a benchmark result. It is a real engineering project that previously required a full team and a multi-month plan. The honesty improvement compounds that: a model 4x less likely to silently pass broken code is one you can trust running overnight without supervision. For anyone building on Claude Code today, this release turns it from a coding assistant into a distributed engineering system.

OpenAI shipped Codex version 26.527, adding Computer Use to Windows for the first time and remote control from ChatGPT on iOS and Android, allowing the coding agent to see, click, and type inside any Windows application directly from the active desktop. This brings a capability that previously existed only on Mac to the platform where most enterprise software actually gets built.

For a year, Codex lived inside the editor: completing functions, explaining stack traces, generating tests. Windows Computer Use expands the surface to the full desktop. The agent can now verify a UI bug that only appears after clicking through three screens, reproduce a flaky installer, or sort a Downloads folder by file type in under three minutes, all tasks that previously required a human physically at the machine. The constraint is that Codex runs on the active foreground desktop, so the machine becomes the task surface while it works. But this is the first time an AI coding agent has had real hands on the platform where most enterprise software gets built, tested, and shipped.

Microsoft will unveil a suite of homegrown MAI models at Build 2026 in San Francisco, including a dedicated coding model built to take GitHub Copilot back from Claude Code, per reporting from Reuters and The Information.

The lineup includes MAI-Image-2.5 (currently ranked #3 on the global Arena text-to-image leaderboard behind only OpenAI and Google, with image editing support and a faster "2.5e" variant), MAI-Transcribe-1.5 (speech-to-text, lowest word error rate across 25 languages), and MAI-Voice-2 (multilingual text-to-speech)

The takeaway for builders is about the shape of the distribution, not a doomsday headline. If AI is better than average at structured creative tasks, the work most exposed is the kind that needs only average creativity, while genuinely exceptional or culturally rooted work stays protected longer. The practical move is to use AI to lift the floor on routine creative output and spend your own time where real originality is the differentiator.

A new all-in-one AI workspace lets you spin up a personal agent that runs on its own always-on cloud computer, accesses every major AI model, and executes multi-step work across your apps, so you don't have to pay for a dozen separate AI tools anymore.

Head over to Genspark and create a free account. Once you're in, you'll land on a dashboard full of "super agents", each one purpose-built for a different kind of work. You'll also notice you can register for free credits to test everything before committing to a plan.

From the dashboard, select Genspark Claw to create your first AI employee. This is the part that makes it different from a regular chatbot: instead of typing a prompt and getting text back, you message Claw like a coworker, and it actually executes every step for you.

Each Claw runs inside its own dedicated, always-on cloud computer, so the work keeps going even after you close your laptop. Try something like:

"Research the top 5 competitors in my space, summarize each in three bullet points, and send me a Google Doc."

Claw will handle the whole thing end to end and come back with the finished result.

To trigger your agent without leaving whatever page you're on, install the Chrome Extension. It sits inside your browser as a sidebar that understands the context of any webpage you're looking at.

This means you can highlight something, ask a question, or kick off a task right from the site you're browsing, no copy-pasting back and forth into a separate tab.

For calls, enable Meeting Bots. Connect your Google or Outlook calendar, and a dedicated bot will automatically join your scheduled meetings, take notes throughout, and send back a full summary with action items, before you've even closed the tab.

Pro Tip: Because every model lives under one roof here, you can route different tasks to different AI models depending on what you need, research, writing, image generation, or coding, without paying for separate subscriptions for each one.

At the start of this year, I did an audit of my week.

I wrote down every task I do on a daily and weekly basis. Scripting videos. Writing titles. Researching markets. Creating proposals. Drafting messages. Generating content ideas.

Then I calculated how many of those hours I was spending just prompting AI to do those things.

10 hours a week.

10 hours of opening new chats, explaining context, iterating, giving feedback, and getting outputs that were almost right but not quite, only to start the whole thing from scratch the next day.

That's when I realized: the prompting itself is the problem.

Skills are the solution.

You put in the work once. You build the skill. And from that point on, every output you ever need from that workflow takes one command, a single forward slash, and you get exactly what you want, in exactly your format, every time.

This guide covers three methods to build them. I use all three.

Claude has a built-in skill that builds other skills. Most people don't know it exists.

After that, any time you want to use it: /your_skill_name {your request here}

The scheduling upgrade:

Once you have a skill, you can schedule it. In Cowork, go to Scheduled → New Task. Then write something like:

“Use my market brief skill daily at 8:00 AM to create a crypto and market report.”

Gold:

“Generate video ideas for me every morning based on my video script skill.”

Your skill now runs on autopilot. You wake up and the output is already there.

Become a paying subscriber to get access to all premium content.

✓ Full archive of premium guides with ready-to-use prompts

✓ Structured AI courses (step-by-step, start-to-finish)

✓ Every upcoming premium tutorial

NVIDIA launched Cosmos 3 at GTC Taipei on June 1, the world's first fully open omnimodel for physical AI combining vision reasoning, world simulation, and action generation in a single model. Cosmos 3 Nano (8B) and Super (32B) are downloadable on Hugging Face now under a commercial-use license.

ElevenLabs released Music v2 on May 31, an AI music model that switches genres mid-track, lets you regenerate individual song sections via prompts without touching the rest, and embeds sound effects directly in compositions. Trained on fully licensed data, with API pricing cut up to 50%.

Anthropic confirmed in its Opus 4.8 launch materials that Claude Mythos Preview, currently restricted to around 50 Project Glasswing partners, will be publicly available "in the coming weeks." Mythos identified 23,019 security vulnerabilities in open-source code in its first month of operation.

🤖 Claude Opus 4.8: Anthropic's new flagship, just released. Best coding benchmarks available right now, 4x more honest about bugs, and fast mode is 3x cheaper.

🪟 Codex for Windows: OpenAI's coding agent just landed on Windows with full Computer Use. It can see, click, and type in your desktop apps.

🎵 ElevenLabs Music v2: Genre-switching AI music generator with section-level editing. Fully licensed for commercial use. API pricing just dropped 50%.

Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!

See you tomorrow :)

- Dr. Alvaro Cintas