I run an always-on AI team on my Mac mini, my own little Q branch. Mane, the team lead, splits each request into jobs and briefs everyone else on a need-to-know basis. Every agent plays one of three roles. Some know a lot about one thing, some do the work, and some check that it was done well.
And cheap subagents just got seriously capable. Anthropic launched Haiku 5.5 on October 7 and pitches it as a fast, low-cost subagent that a bigger model hands work to. My rule of thumb now: Opus or Astra to plan and orchestrate, Sonnet or Sol for most subagent tasks, and Haiku or Luna for grunt work like moving files, renaming them or fetching information.
Silent 100-second animation: how my AI team puts a website live on Cloudflare Workers.
This week, agents got easier to use and more proactive than chat ever was. And Cowork went cloud-first, which changes where your local work should live.
What happened: At DevDay on September 29, OpenAI introduced Dots: personal agents that keep your context, preferences and ongoing work. Each dot has its own cloud computer, so it keeps working between conversations. When you're not talking to it, it does what OpenAI calls "proactive research," looking for ways to help in the apps you've connected.
You start with one dot. OpenAI says it "envisions" teams of dots, but they aren't here yet. Dots are rolling out on Pro and Business Premium in eligible markets (here are the plan, region and usage limits). Custom Rules set what your dot is licensed to do: allow an action, require approval or block it.
It isn't alone. Grok Bot already runs a long-lived assistant on a persistent cloud computer, and Meta's Muse flags emails and prepares drafts for small businesses, with approval required for sending, publishing and spending.
What changes: Chat waits for you to ask. These agents keep going after you leave and come back with work.
Orchestrators like Mane still have an edge: more consistent, more thorough memory, and true subagent orchestration through repeatable skills and agent personas.
Two cautions. OpenAI's own advice: "Dots can still make mistakes, so always review consequential work." And check access: Grok Bots have separate memories but share one account computer, browser sessions and app logins included.
What to do next: Give your dot one area of your business to watch, like a key client or a product launch, and ask it to keep you up to date. Then connect what it needs to see: your meetings, email, Jira tickets. OpenAI says dots can connect to over 4,000 apps, and their background research uses read-only tools, so they can watch without sending or changing anything. Add a Custom Rule so it asks before acting on what it finds.
What happened: As of October 6, new Cowork tasks on Pro and Max run in the cloud, and the local-only setting is gone. New scheduled tasks run in the cloud too. Scheduled tasks you already have stay local. Anthropic's support guide has the details, and admins can see how each mode is isolated in the architecture documentation. (On Team and Enterprise, cloud Cowork is in beta.)
What changes: Cowork has basically merged with chat into a hybrid. It has some memory, and it still reaches your tools and systems. It's excellent when everything you need comes through a built-in connector or a hosted MCP server. (Local MCP servers don't run in cloud sessions.)
What it gives up is the power of a local system. Each cloud session runs in a temporary sandbox that's destroyed when the session ends, so it can't write and run code repeatably. It reaches your local folders only while the desktop app is open. And there are size limits on the files and GitHub repos you can share with it.
That's why Mane runs in Claude Code on my Mac mini, where it works directly with local files and runs the same code again and again.
What to do next: List the Cowork tasks that touch local files, large sets of files or code you want to rerun. Move what fits to a connector or hosted MCP. For the rest, switch to Claude Code. Anthropic's guide shows how to bring a task over with "Download task data" and /port-task-data.
That's it! Which recurring job would you give an agent, and where should it stop and check with you? Hit reply and tell me. Replies are for my eyes only.
–KM
P.S. How many James Bond references did you find?
|
WORK WITH ME I'm Kevin. At DianoAI, I help companies build agent teams: shared agents with IT and personal agents employees can improve. If you're trying to do that, reply and tell me where you're stuck. We'll see if working together makes sense. |
A weekly briefing for business leaders who want the few AI developments that matter.