Hello!
I just wrote about an overhaul of the content indexing system behind Ask LukeW (the thing you're using right now). The system now checks hourly for new or changed content and runs audio, video, PDFs, and images through an ingestion pipeline without anyone having to kick it off manually. Audio and video get transcribed with speaker identification, and PDFs go through a vision model that describes charts and diagrams. That last part matters a lot, since my slides are mostly visuals. The biggest addition? Nearly 20,000 of my X (Twitter) posts. New ones come in automatically via webhook (a notification sent the moment something gets posted), while the older ones came through a one-time import. Each post also gets enriched with expanded links, quoted post context, and AI descriptions of any media. Why bother with all this plumbing? Because I keep relearning the same lesson. The quality of the data going into an AI system drives the quality of the results coming out. When we first added about 4,000 PDFs, the answer to "what is design" turned into something like "design is the fundamental soul of a human made creation." I'm still not sure who says that, but it isn't me. Junk layout markup was crowding out everything else, and only cleanup and later switching from text extraction to vision model descriptions fixed it. What's new this time is that the pipeline runs on its own. A richer, always-current corpus also made a new feature possible: the experience now starts with an answer instead of waiting for a question, so people can see what the tool can actually do. As I've said before, adding a chatbot doesn't mean you're done. That's the beginning. Most of the real work is in keeping the content flowing in clean.
How are AI interfaces evolving beyond chat toward agents?
In just a few years, AI products have moved through four distinct stages: models behind the scenes powering features, then people talking directly to models (chat), then models deciding which tools to use and how (agents), and now agents orchestrating other agents. Each shift has changed what the interface needs to do.
Before chat got all the attention, deep learning was quietly improving Google Translate and YouTube recommendations back in 2016. Then chat put models front and center. But anyone who's used a chat AI knows responses quickly get lost in conversation threads, and producing something from a pile of chat replies can be painful. So many apps added a canvas or artifact pane (a dedicated area for the output), splitting the screen roughly half input and half output.
Agents pushed chat back even further. Instead of you steering every step, people state their objectives and the model plans which tools to use and how to accomplish the task. You can see this in Bench: it went from a single-pane chat UI to a split pane, and now chat takes up less than a fourth of the screen and is collapsed by default. Your job becomes kicking things off, clarifying when needed, and making use of the final output.
Which brings us to today's design challenge: orchestration. You manage agents, they manage more agents, and those agents manage even more agents. Chat threads break at that scale. As anyone who's tried to manage a team through a single Slack or email thread can attest, one conversation doesn't cut it. In Intent, we set things up so you define what you want to get done and let specialized agents work in parallel while staying aligned.
That also means making the agents' work visible. Showing research, planning, tool use, and generation steps as they happen helps people trust what's going on and know when to step in.
And inboxes and dashboards aren't enough on their own. They answer "what are my agents doing right now?" but more specific and dynamic views may actually be more useful. That's why Intent lets you create any number of notes (documents) per task, so you get the view you need when you need it.
Put simply, the design question has shifted from how people interact with AI to where and how people fit in. If you're designing AI products today, spend less energy polishing the chat box and more on how people delegate, review, and steer.