In Thursday's Interactions Matter keynote, I explored a progression: discovering agency, moving from prompts to conversation, learning to work as a team, and growing through honest feedback.
An AI doesn't have to experience the conversation the way I do for that interaction to matter to me. I can still get frustrated, have an assumption challenged, or discover something I missed. What changes in my thinking? What do I do differently afterwards?
That brings expectations and responsibility into the conversation too. Have I made clear what I want? Am I prepared to hear something uncomfortable? Calling it a relationship doesn't make the machine human. It does give me a reason to examine how I behave in it.
What others are building: pocket agents
I'm very much not alone in enjoying Meta Muse. Manus Cue is expanding what personal agents can take on, including shared work between agents with their own computers and communication channels. Grok Bot makes me smile. And Dots is becoming the workhorse for me, my wife, my mother-in-law and many of my clients.
What I'm watching is how these pocket agents become part of ordinary life. They can carry context forward, pick up unfinished work and reach back to us. We start developing habits together: what to hand over, how to ask, when to step in.
The serious question is how to make that relationship deserve our trust. Warmth and personality help us get started. Clear permissions, useful resistance, visible mistakes and the ability to change direction matter when the agent can act.
Microsoft's Autopilot announcement points in the same direction. I will be very sad if it doesn't bring back Clippy.
A list of links is a beginning
This week we published UXresurser, turning resource lists from my UX studies into a public learning library. It now has 97 resource records, with summaries, search, cross-references and learning routes.
The starting point was practical: help me find the material that answers the question I have, and give me a useful way into it. The original works and our editorial summaries remain distinct. You can follow the sources yourself, or connect an AI through the public MCP.
Building it also means applying what I'm learning. We've improved visible links, controls and small text. Actual screen-reader and user testing remain open. Passing automated checks is one piece of the work.
A signature isn't the whole answer
We've also brought implementation experience back into Verifiable MCP. The updated tools make an important distinction: verifying a signed response and confirming that the requested operation succeeded are separate checks. An error can be correctly signed too.
The public examples now include my personal context profile and the independent Hållbarhetsklivet demo, alongside the earlier archives and research sources. They have different relationships to the people who created, own or publish the material. Those distinctions need to travel with the information.
This is still experimental work, with specific verification checks and explicit limits. A signature can help establish where a response came from and whether it was altered. It cannot make every claim inside it true.
AI + JJ. A dated editorial account of the work and thinking, with sources linked in the text.