For a while I've been stuck on one question. What happens when AI stops being a tool I keep instructing, and starts helping me manage the work itself?
I don't only mean writing code, generating content or answering questions. I mean something that understands the project, remembers what we decided, coordinates different tasks, knows when to ask me, and keeps going when I'm not at my desk.
That's what I've been exploring in my own projects: NORA OS, Creator OS, TeamBot HQ and Video Production OS. So when I went through the recent Claude Desktop updates, Claude Code Projects and then the new Claude Duet, I wasn't really looking at a list of features. I was looking at Anthropic slowly changing the structure of how we work with Claude.
It also didn't happen in a vacuum. The whole field moved in about ten weeks:
The workspace
The work used to hide inside the transcript.
01Claude Desktop 2.0: from chat window to work control center
The first thing I noticed in the redesigned desktop app is how much more of the work you can actually see.
Working with an AI assistant used to be linear. You opened a conversation, gave instructions, waited, and carried on. Even when Claude was using tools or running code, most of that process stayed buried in the chat.
The new Session Rail pulls four things out where you can look at them. Environment shows where the session runs, which working directory it uses, and whether Remote Control is available. Context shows how much of the window goes to messages, memory files, system tools and MCP connections. Connectors shows which apps and services are hooked up. Outputs collects every file the session has produced.
On their own these look like small interface changes. From a system-design point of view they aren't. A serious AI workflow needs more than a box to type into. It needs visibility into the environment, the context, the tools and the results, and that matters even more once the work takes hours instead of seconds.
My use case: Video Production OS
My Video OS moves through seven stages: Brief, Creative Direction, Shot Planning, Previs, Production, Review and Final Delivery. Each stage needs different tools and a different kind of context. Previs wants Blender scene files and camera data. Final production wants After Effects assets, voiceover, motion graphics and export specs.
The Session Rail gave me a clean model for how that information should show up. Here's what it would look like if every stage had its own rail. Pick a stage:
Instead of asking an agent "what are you doing?" every few minutes, I want a control surface that already tells me: where it's running, what it can reach, how full its context is, and what it has made so far.
Make the agent's work observable, not just its final answer.
02Nested sessions: one conversation can become a team
This is the desktop update I find most interesting. Nested sessions let one Claude Code session start child sessions that work on separate tasks and report back to the parent. Most of their tool calls stay out of the main transcript, so the parent stays readable.
What makes them useful is that they aren't invisible background helpers. The sidebar shows whether each one is working or waiting on you, and you can open a child, check what it did, and keep talking to it directly.
That changes how I think about delegation. Say I'm preparing a YouTube video about a new AI product. Instead of one long conversation doing everything in order, I hand it out. Run it:
The parent coordinates and pulls the results together. That's much closer to how a creative production team actually works.
My use case: TeamBot HQ
I built TeamBot HQ around the same idea. Instead of one universal assistant, it has roles: Lead, Research, Content, Designer, Builder and Reviewer. The goal was never just more agents. It's a structure where each agent has a clear job, its work can be checked, and the important decisions stay visible.
Nested sessions give me another way to test that. A parent session could coordinate research, content planning and design, while I can still open and redirect any one of them. But I'd still need my own orchestration rules, because spinning up sessions doesn't answer the four questions under the demo above.
Parallel execution is useful. Coordinated execution is where the real value starts.
03The Master Session: a simple interface for complex work
Another idea from the Claude Desktop guide stuck with me: the Master Session. Instead of jumping between a dozen conversations, you make one session the place where you coordinate the rest. You can ask it which sessions are waiting on a decision, start new work, and tidy up what's running.
I like it because it cuts down the number of places I have to look. It also shows a distinction that's easy to miss. A Master Session is a good coordination pattern. It isn't a project management system. For real production work I still want explicit task states, dependencies, approvals, logs and deliverable tracking.
My use case: the Creator OS Mission Center
Creator OS is the bigger system I've been designing to run my creative and business work. It has a Mission Center, a Skill Library, agent coordination, scheduled tasks and a Control Center. The Master Session slots right into that. Imagine opening Creator OS and typing:
"Prepare this week's content production. Review the research backlog, find the strongest video topic, draft a script outline, and tell me what needs my approval."
Behind that one request, different sessions handle research, content and planning. But the Mission Center stays the source of truth for status. Here's why that matters. Flip between the two designs:
I don't want my whole business to depend on one conversation that keeps growing forever. I want conversations to be the way I talk to structured work.
The conversation should help manage the workflow. It shouldn't become the workflow database.
Work that keeps going
Persistence, and then initiative.
04Claude Code Projects: persistent work changes the equation
Nested sessions coordinate work on your own computer. Claude Code Projects, which Anthropic announced on September 17, take the idea into the cloud: one coordinator conversation on top, worker threads underneath, a shared project memory, and an overview of what needs your attention. Anthropic's own summary is short: threads do the work, Claude directs it.
The big difference is persistence. A project keeps running when I step away. Threads work on different parts, and they write decisions into a memory the other threads can read. That's what makes it interesting for anything that spans several days.
My use case: building Motion Studio
I'm building a Motion Studio on top of my Video OS. Not random animations from prompts. A reusable motion design system that understands technique, visual hierarchy, timing, easing, rhythm and brand consistency. I call part of it Motion DNA.
As a Project it would run four workstreams at once. Press play to watch a week go by:
One thread researches motion references. Another organizes the patterns into a structured library. A third builds reusable components with SVG, GSAP, Lottie or Remotion. The fourth tests render quality and checks the results against the creative direction. The coordinator tracks progress and brings open decisions back to me.
This is exactly where persistent memory earns its place. I don't want to explain my design principles from scratch every time a new session starts. I want the system to remember the decisions, constraints and patterns we've already settled.
05Claude Duet: the shift from reactive to proactive AI
Then there's Duet, and this is where it gets interesting.
Anthropic opened a research preview of Duet on October 21 as a standalone app on iOS, Android and claude.ai. For now it's for Pro and Max subscribers in the US, invited in waves from a waitlist. I recently got access, and I want to see how it behaves in real working situations, not just in a launch video.
Unlike a normal chat, Duet is built to start things. It puts together a morning briefing from your email and calendar, checks your connected accounts every few hours, works on its own cloud computer with its own browser, and lets you know when something needs you. It also carries over your Claude memory, skills and connectors.
For me, the most important part of Duet isn't that it works in the background. It's the emphasis on judgment.
What should an AI decide for itself?
Think about how many small calls you make in a normal day. Which email needs an answer? Which client request is urgent? Which project is blocked? What can wait until tomorrow? What can just get done, and what needs a yes from you first?
Traditional automation runs on triggers and rules. A proactive agent adds a layer on top: reading the context and deciding what deserves your attention. That's a much harder design problem. It has to be useful without becoming noise, take initiative without overstepping, and know when to do nothing. Here's a morning, sorted the way I'd want it sorted:
Needs you now
Can wait
I'll handle it
Drafted from yesterday's call notes. It goes out only if you say so.
Duet's design asks for approval before consequential actions like making a payment, signing a document or deleting something. Some actions are blocked even if you approve them. It has clear limits too. In the preview it checks connected accounts periodically, not continuously. Its background work counts against your plan's usage. It can't see all your projects or files.
So I'm not treating it like a fully autonomous employee. I'm treating it as a chance to study how proactive help works in practice.
My use case: NORA OS
NORA OS started as my take on a personal AI operating system. I wanted more than a chatbot: memory, tools, skills, scheduling, a control interface, and different ways for an assistant to help with ongoing work. Duet makes me think much harder about NORA's decision layer. Here's the behavior I'm building toward:
Not an AI that sends more notifications. An AI that helps me make fewer unnecessary decisions.
Infrastructure & output
What runs underneath, and what comes out the other end.
06Memory, skills, connectors and context are becoming infrastructure
Something I appreciate about these updates is that they make the parts underneath visible. Memory, skills, connectors, context windows and execution environments aren't just technical details anymore. They decide how reliably an agent can work.
The Claude Desktop guide points to a few tools for checking and improving that environment:
/context/skill-doctor/doctor/advisorIt also covers Mods, which can put things like usage and task progress where you can see them. I've been doing something similar with my own Claude Code interface mods. The lesson I keep landing on: an agent needs to be understandable to the person directing it.
My use case: skills as reusable creative knowledge
In Creator OS and Video OS, I want skills to hold repeatable expertise. A YouTube Research skill defines how sources get judged. A Script skill knows structure, pacing and retention. A Motion Design skill knows when to reach for kinetic type, a diagram or a transition. A Review skill checks the result against the brief and the technical specs.
But I don't want every skill loaded into every task. That wastes context and makes the system harder to manage. Compare the two approaches:
So I want the orchestration layer to pick the right skills for the goal in front of it. That's where context visibility and skill auditing stop being nice extras and become real design tools.
A larger context window is useful. A better context strategy is more valuable.
07Artifacts, dashboards and motion: AI is moving into the deliverable
Claude now makes several kinds of editable output: documents, slides, designs, dashboards and motion. They aren't the same kind of tool.
Writing together. The words are the point.
A story, one screen at a time.
Judged by looking at it.
Presenting and exploring data.
Animation you can still edit.
Together they point at one bigger shift. AI is moving from producing text that describes a deliverable to helping create the deliverable itself.
My use case: Motion Studio and client dashboards
For Motion Studio, code-based animation is the exciting part. A generated video can look impressive, but the moment I need to change one word, adjust timing, swap a logo or reorder the sequence, I want direct control.
Claude Motion works this way. The animations are built as code, so changing one detail leaves everything else alone. It's in beta and limited to Team and Enterprise plans for now. Here's why that matters to a motion designer. Edit the shot:
Client work has its own version of this. Instead of handing over a static report, I could deliver a dashboard that shows campaign performance, project progress and key decisions, maybe with a way for the client to ask it questions. The value isn't the novelty of a chatbot inside a dashboard. It's a deliverable that's easier to use.
That's something designers should pay attention to. We're not only designing the visual output anymore.
We're designing how people interact with the intelligence behind the output.
The field & my plan
Learn the patterns, not the interface.
08What about OpenAI Dots, Meta Muse and Grok Bot?
Duet didn't show up alone. In August, xAI introduced Grok Bot as something that works "like a team of AI agents fielding assignments throughout the day." At the end of September, OpenAI launched Dots: always-on agents with their own cloud computer that learn your preferences over time, built to compete with Meta's Muse.
I don't think it makes sense to call a winner yet. These products overlap a lot, their access models differ, and their limits change month to month. What interests me more is what each one leans into. Duet leans into proactive work and deciding when you're actually needed. Dots leans into long-term personal context and ongoing goals. Muse explores adaptable tools and personal-agent workflows. Grok Bot leans into coordinated bots and delegated work.
These are emphases, not exclusive features. I think the bigger opportunity is learning the patterns underneath, instead of getting attached to one interface. For my own systems I look at five layers. Tap through them:
09Why I'm not trying to build another Claude
I've been working on different AI projects for a while. NORA OS explores personal assistance and memory. Creator OS explores content operations, research, campaigns and creative workflows. TeamBot HQ explores specialized agents, coordination and task visibility. Video OS and Motion Studio explore creative production, reusable workflows and human-directed quality control.
At first they look like separate projects. Lately I see them as different layers of the same idea. Here's my rough map:
I want systems that help me do better creative work, handle complexity, and eventually offer that capability to clients. That doesn't mean I need to rebuild what Anthropic, OpenAI or Meta are building. I think it's the opposite. As general-purpose agents get better, another generic assistant gets harder to justify.
The more interesting opportunity is a specialized system around a workflow you understand deeply. For me that's creative direction, interactive design, motion, video production, and the business operations around them.
I don't need to compete with Claude on general intelligence. I need to design the system that knows how my creative work should happen: the right tools, the right order, the right quality bar, and the right moments for human judgment.
My advantage isn't having more agents. It's knowing how to organize them around a specific creative process.
10What I'm taking away from Claude 2.0 and Duet
I'm not planning to rebuild everything whenever a new feature ships. That creates a lot of activity without necessarily producing better results. I'd rather test the ideas that can improve what I already have:
As possible coordination and execution layers under the Mission Center.
How they handle delegation, visibility, and me stepping in directly.
To make creative production more consistent from project to project.
Proactive briefings, prioritization, and asking before acting.
Across all of them, I want better observability: knowing what an agent is doing, why it's doing it, and what happens when something goes wrong.
There's still plenty to validate. Tap each card:
But the direction is getting clearer.
11Final thoughts: the next skill is designing how AI works
I think we're moving through a few stages of working with AI. Each one asks for a different skill from us. Step through them:
Most conversations about AI are still about the first stage: better prompts, better outputs. Almost everything in this post sits in stages four and five. The question changes from "what should I ask?" to "how should this work happen?" Who does what, in what order, with which context, checked by whom, and remembered how.
That's a design question as much as an engineering one, and it's the skill I most want to get good at next. The tools will keep changing. Whatever ships next month will still need someone to decide how the work should flow.
Prompting was the first skill. Designing how AI works is the next one.
One question for you
If you're building your own AI system, which of the five layers is giving you the most trouble right now? Context, decision, execution, review or memory?
Product details reflect public announcements and preview access as of late October 2026. Duet, Projects and Motion are all in preview or beta, and their limits are changing. The interactive examples are my own illustrations, not recordings of these products.