02 / Pocket agents and AI teams
What would your
Marvin do?
The interesting question is what help would matter in your life.
A useful connection.
I showed a small example from my wife Kattis's life. Her agent, Marvin, connected information about an expected delivery with her plans and noticed a mismatch. It brought the issue to her attention on her watch, where it was easy for her to notice and act.
The point was the connection between context, timing and a useful next action. The individual pieces of information were not remarkable. Putting them together in the right moment made them valuable.
I also mentioned Stephen, the agent my mother-in-law interacts with, and the assistants becoming part of my own working day. These are personal examples of a possibility to explore. Your version may look different.
What would you want an assistant to notice before you had to ask?
Explore my own assistant
Help me imagine one useful role for an AI assistant in my everyday work or life. Ask about the recurring things I forget, coordinate, check or worry about. Ask one question at a time. Choose one concrete situation. Explain what information the assistant would need, what it could propose, what it should ask me before doing, and how I would know it had made a mistake. Begin with the smallest useful version using the access and tools I actually have.
A conversation to keep having
Curiosity, with judgement.
Does this make sense to me?
Which part feels useful, exciting or uncomfortable? Name the specific situation behind that feeling.
What would it do?
Choose one job. “Help with everything” is harder to evaluate than “notice when two plans conflict”.
What should it know and be allowed to do?
Reading a plan, drafting a reply and sending it are different permissions. Give each a deliberate boundary.
What could go wrong?
Consider mistaken context, unnecessary disclosure, dependence and actions taken too early. Decide how you would notice, stop and correct them.
What is the first experiment?
Make it small enough to inspect. Keep asking what is improving and what needs to change.
The same approach helps with broader questions about privacy, resources and the social effects of AI. Take them seriously, look for sources and distinguish claims about the whole industry from the consequences of a particular use. A useful question does not need a predetermined optimistic or pessimistic answer.
A team, rather than one perfect tool.
I showed my AI-team catalogue to explain how different services can contribute different capabilities. Research, transcription, writing and production do not always need the same tool.
The catalogue and its costs describe my own setup at a point in time. They are not a shopping list, a minimum requirement or a claim that each service is a continuously autonomous colleague. You can begin with one tool you already have.
In the questions after the talk, I described a coordinating agent that breaks work into tasks and specialists that pass their outputs through a shared working space. A finished research note can become the input for the next task. Clear handovers make that cooperation inspectable.
One clear outcome.
The coordinator needs to know what should be better when the work is finished.
Specialists with useful context.
Each task needs the relevant sources, constraints and definition of a good result.
Work that can be handed over.
Save sources, decisions and outputs where the next contributor can find and check them.
Someone remains responsible.
More agents create more coordination to get right. Keep decisions, permissions and checks visible.
Watch the direction.
Test the actual capability.
I used Muse from Meta as one pocket-agent example. Its current Apple listing and Android listing are linked for further exploration. The examples illustrated a direction: assistance that can stay with a task, retain useful context and bring work back to you. Product names, availability, ratings and access conditions move quickly. I have not turned those on-screen snapshots into a buying guide here.
Use the assistant you can actually access. Ask it to explain its current capabilities, then test them on a small task. A convincing description of what it can do is not the same as a completed action.
Read with your AI · Markdown →
About this companion and its sources
Prepared from the final 33-page presentation, the Circleback session transcript and selected frames from its screen-share recording. The pages follow the delivered talk and expand its methods into practical exercises. They are an edited companion, not a verbatim transcript.
Karin, Marvin and my working setup are my accounts of experience. The Kasparov framework is my teaching interpretation. The fear curve is an illustration. The 16-hour learning plan is a target-based example, not measured savings. Product snapshots and broad statistical claims have not been republished as current facts.
The recorded closing questions are paraphrased without attendee names. Meeting-chat contributions and full poll results will be added after they have been supplied and reviewed. The raw recording, private context and source transcript are not published here.
AI + JJ · 12 October 2026. No added analytics, sign-in or data collection. The appearance preference is stored only in your browser. External links have their own services and policies.