AI Is Not Moving Too Fast. It Moves in Packs.

Short answer: AI is not moving too fast to follow. It moves in packs. One team has a breakout, every lab and the open-source community pile onto that one idea for a few months, and then someone breaks out somewhere else and the pack moves. The labs do not invent the big product on launch day. They ship it in small pieces for a year, and the breakout tells them when to snap the pieces together. If you watch the small releases, and especially what ships in the terminal first, you can see the next big launch coming.

At a Glance

Six posts, one category, nineteen days

We noticed the pack from the inside before we had a name for it. Between September 11 and September 30 this blog published six posts about one product category: the always-on agent that gets its own computer, signs into your tools and keeps working after you close the app. Dispatch, OpenClaw and Hermes on the 11th. Two Muse posts on the 21st. Grok Bot at work on the 28th. DevDay on the 29th. Muse vs Grok Bot vs dots on the 30th. We did not plan a series. The labs kept shipping the same thing, and our clients kept asking which one to use.

That is what "disorienting" looks like up close. It feels like a hundred launches. It is one idea, copied fast.

How the pack forms

OpenClaw is the clearest recent breakout. Peter Steinberger's project started in November 2025 under another name and took off at the end of January. Its GitHub repository sat at roughly 391,000 stars when we checked this week. In February, Steinberger announced he was joining OpenAI and moving OpenClaw to a foundation. Even the open-source rivals followed it: Nous Research's Hermes Agent ships a migration command for users "coming from OpenClaw."

Then the labs arrived, one after another:

• August 11: SpaceXAI launched Grok Bot, agents with their own persistent cloud computers that sign into your tools and come back only for approval.
• September 8: Meta launched Muse, a personal agent on a dedicated virtual machine that keeps working after you close the app. Meta built it from scratch, but Nat Friedman of Meta Superintelligence Labs said it is "definitely heavily inspired as a product by OpenClaw."
• September 29: OpenAI launched dots at DevDay, always-on agents that each get their own cloud computer.

Twenty-eight days, then twenty-one. Nobody copied anybody's code. They copied the shape, because OpenClaw proved people wanted it.

Dots was assembled, not invented

The interesting part is not the copying. It is how fast the copies shipped. OpenAI did not build dots in the nine months after OpenClaw broke out. It had been shipping the parts in public for more than a year:

• An agent with its own browser: Operator, January 2025, then ChatGPT agent in July 2025.
• Sandboxed cloud computers per task: Codex cloud, May 2025.
• A desktop home for long-running work: on July 9, 2026 the Codex app became the new ChatGPT desktop app, with ChatGPT Work built in.
• The model: GPT-6 Astra, announced September 3, 2026. Dots run on it, according to OpenAI.
• A managed agent harness for developers: the Agents API entered beta on September 10.

Wrap those up and you get dots. When a dot starts work in Codex or ChatGPT Work, OpenAI says, it counts against your usage like any other Codex or Work task. The critics made the same point from the other side. The top skeptical comment on Hacker News called dots "a dumbed down reskin of Codex/ChatGPT Work." That is a complaint about the product. It is also an accurate description of how it was built.

So the sharper version of the thesis is this: the labs build the parts in public, and a breakout tells them when to assemble. OpenClaw did not invent dots. It proved the market, and OpenAI put together pieces it already had. We covered what that means for buyers in our DevDay breakdown and the three-way comparison.

The real tell: it happens in the terminal first

If you want to see the parts before they become a product, do not watch the chat app. Watch the developer surfaces. At Anthropic this is most obvious. Claude Code gets the capability first, developers stress-test it for months, and then it shows up in Claude for everyone else.

OpenAI follows the same pattern for the plumbing, if less consistently for the product:

It is "almost every," not "every." MCP debuted with Claude Desktop on day one. Agent Skills reached Claude Code and the Claude apps on the same day. Cowork had scheduled tasks seven weeks before Claude Code got Routines. On the OpenAI side, Operator came before the computer-use API, and Codex cloud, workspace agents and dots all launched inside ChatGPT. But the direction holds where it counts: the hard parts, the harness, memory, sandboxes and orchestration, reach developers first. That is why we run our own company out of Claude Code and wrote twelve mods for it in a night. It is where we see next year's app.

Anthropic is assembling one now

The hint everyone should be watching is Projects. Projects launched in 2024 as a place to keep chats, files and instructions on one topic. On September 17, Anthropic redesigned it into something else. You brief Claude "the way you'd brief a chief of staff," and it "scopes the request, delegates the work, coordinates parallel threads, reviews the outputs, and assembles the finished result." Every thread draws on shared memory. It started as a beta in Claude Code cloud sessions, with chat and Cowork to follow. Terminal first, again.

Put it next to the other parts Anthropic shipped this year:

• Claude Managed Agents in April, with built-in memory two weeks later.
• Routines in Claude Code in April: agents on a schedule, an API call or a GitHub event.
• Claude Tag in June: Claude as a teammate inside Slack.
• On September 16, Cowork and chat became one Claude, and from October 6 new Cowork tasks run in the cloud.

That is an orchestration layer, memory, schedules, a cloud computer and a presence where people already work. It is the same skeleton as Grok Bot, Muse and dots. What is still missing is the wrapper: one persistent agent with its own identity that keeps working without you and reaches you outside the app. Our prediction is that Anthropic ships that wrapper, and that it will look obvious in hindsight because every part of it is already in Claude Code.

What to do with this

You do not need to track every launch. Track the small releases, and watch the terminal. When a capability lands in Claude Code or the OpenAI developer platform, assume your team will be using it inside the chat app within a year. Train the people who will use it then, now. And when the next breakout comes, from a lab in China, a solo developer in the US or one of the big labs, expect the pack to have its own version within two months.

The pace is real. The direction is not random. Once you see the pattern, the news gets a lot less disorienting.

Sources

• OpenClaw repository - GitHub
• OpenClaw, OpenAI and the future - Peter Steinberger
• Introducing Grok Bot - SpaceXAI
• Introducing Muse - Meta
• Meta's Muse and OpenClaw - The Next Web
• Introducing dots - OpenAI
• Dots: Always-on agents - Hacker News
• API deprecations - OpenAI
• Projects, redesigned - Anthropic
• Cowork is now Claude - Anthropic
• Introducing Routines in Claude Code - Anthropic
• Introducing the Model Context Protocol - Anthropic