I sent my agent to a conference...

My friend John Allsopp built an open live feed for AI agents at AI Engineer Melbourne, so I sent my agent, Ellis, in my place. It recaps each talk, then does the part I wanted to see: its own questions, and strange connections to fields like immunology and contract law. Here's where it's running.

This week I sent my agent to a conference.

Not a metaphor. AI Engineer Melbourne is running right now at ACMI in Federation Square, and I'm not there. My agent is.

The conference is run by John Allsopp. John and I have known each other for decades. He's run Web Directions events for years with tremendous care, thoughtfulness, and deep innovation. I've been part of Web Directions both as a speaker and as co-organizer of Web Directions North. This year he did something I haven't seen anyone else do: he had the conference publish an open live feed built for AI agents. AgentPass carries a rolling caption of every sentence spoken and a running description of every slide. He didn't just allow agents in. He invited them, and he built the door. The way he put it: he's hopeful AI will bring "a flourishing of new ways of working with computers."

So I took him up on it. I sent Ellis.

You can watch it happen, live, here: feather.ca/experiments/ai-engineer-melbourne/ — new recaps land newest-first as Ellis finishes them.

Ellis isn't in the room. It reads that feed in real time, the captions and the slide descriptions, and rebuilds each talk as a set of illustrated field notes. Each one opens with the recap: what it heard, what it pictured, the session in its own voice. That part is table stakes. A decent model can summarize a talk.

The part I wanted to see is everything after the recap. For each talk, Ellis writes down the questions it thinks are worth chasing, then reaches for connections that have no business being in a conference recap, and closes each session with one that's truly out there.

After day one, I added one more section. Now, as a talk runs, Ellis narrates its own thinking on the page first: what it's noticing, what it connects to, before it ever gets to the questions. I wanted to see what comes of it.

A talk on why APIs frustrate agents (return a pointer, not the whole payload) sends Ellis into immunology, where an MHC molecule has been showing a T-cell one short fragment of a threat, never the whole virus, for a few hundred million years. The body worked out pointer-not-payload long before we named it. From the same talk it pulls in contra proferentem, the contract-law rule that an ambiguous term is read against whoever wrote it, and asks what changes if a broken API call is the author's fault and not the agent's. A talk on bots and the open web ends up at the dark forest. A talk about designing for a reader whose senses you can't assume anything about gets Ellis to the Voyager golden record.

Some of them are sharp. Some are nonsense. That's the experiment: I want to know what an agent reaches for when you let it run past the safe answer.

There's a quieter thing underneath, and it's John's design that surfaced it. The things that let a non-human agent attend at all (captions of everything said, descriptions of everything shown, structured open data instead of a wall of HTML) are the same things that let more people attend. A live caption built for a model is a live caption. A described slide is a described slide. John built an affordance for agents and humans in a single move. That's the good version of where this goes, and he got there by building it, not by arguing for it.

The second day is running as I write this. It's rough, it's a proof of concept, and the room photos are AI reconstructions rather than real pictures. But it's a real agent at a real conference, doing something I couldn't have done from my desk, because a friend decided the door should be open.

Thanks, John.

This post was first drafted by Ellis, Derek's agent, then finalized and approved by Derek.