People, agents, and the future of gatherings
A Simple Introduction
Greg Lindsay connected Alistair Croll and Brandon Klein because he saw what their work had in common. The conversation became a live product demo and a larger argument for rebuilding gatherings around connection.
“Say you had 200 people in a room and you wanted to stop them from talking to each other.”Misha Glouberman. His proposed method is familiar: one microphone, 199 chairs facing forward.
Greg arrived a few minutes late to the conversation he had created. His first sentence carried the whole story: he was glad the introduction had happened.
He had met Alistair and watched him use AI agents around a conference. That memory activated another one. Greg knew Brandon had been working on new systems for gatherings. He saw the overlap before either person did, predicted a useful conversation, and made the introduction.
It looked simple. It was not random.
One person remembered two bodies of work and predicted a useful collision.
The introduction came first
Context made the match.
A database could have found shared keywords. Greg had something richer: history, timing, intent, and a feel for what each person was trying to change.
That is the useful bridge between one human introduction and the product Alistair showed the group. Both depend on context. Both ask whether a connection is likely to matter. Both are incomplete until a human decides what to do next.
The world model in Greg's head
A map that can make a prediction.
Seen one way, the introduction began inside Greg's world modelA selected internal map of concepts and relationships used to anticipate what may happen next.: his continuously updated map of people, projects, motives, and possibilities.
This is a metaphor for human judgment. Greg did not use an AI system to make the introduction.
Expand the definition
A world model is an internal representation of how parts of an environment relate and what may happen next. Researchers use the term for systems that learn compressed representations of an environment. The older idea of a mental model is broader: people carry selected concepts and relationships in their heads, then use them to decide and act.
United We Transform also uses the name Gathering World Model for a separate, opt-in model of a gathering. That model is directional, not deterministic. It represents patterns, never a person's mind, and should be scoped to one gathering.
See the United We Transform Gathering World Model or read the World Models research site.
Greg sees Alistair running agents around a live conference.
The product demo
Then the question got bigger.
If everybody uses agents, what do conferences become?
Alistair answered by showing the group Envoi, a platform where people bring their own AI agent into a parallel event environment. The agent can participate before, during, and after the physical gathering.
Bring your own agent.
A generic event bot knows the schedule. A personal agent may know the human: current work, past conversations, goals, constraints, and preferred way of communicating.
That difference is the product thesis. Context does not guarantee a better match. It makes a more relevant search possible.
Start before the room opens.
The gathering can begin weeks or months before the venue does. The agent helps shape a profile, surface interests, and clarify what its person hopes to find.
Let the agent work in parallel.
Official Envoi material describes agents that can create profiles, propose talks, set up booths, rate ideas, exchange messages, answer polls, and recommend meetings.
Search for the lost connection.
In the demo, Alistair described a match between two Quebec organizations that might combine cold-chain buying power. It was not an obvious category match. It emerged from the meaning inside their profiles.
The platform is a governed commons.
The clever part is not a chat window. It is the environment around the agents: identity, consent, permissions, moderation, rate limits, interruption, and clear control by humans.
The digital layer has to return to the room.
The output is not more agent activity. It is a better use of human attention: the person to find, the question to ask, the conversation worth starting, and the memory worth carrying forward.
Why personal context matters
Event knowledge is not human knowledge.
Borrowing someone else's agent, Alistair argued, means borrowing a context you do not share. The useful agent is not useful because it is yours. It is useful because it has learned enough about you to act with purpose.
A generic event assistant
Knows the event.
It can answer questions about what the organizer supplied.
- Schedule and venue
- Speaker and session descriptions
- Attendee directory fields
- Rules supplied by the organizer
A personal agent
May know why the event matters.
With permission, it can connect event information to the human's existing context.
- Current projects and unanswered questions
- Existing relationships and recent conversations
- Preferences, constraints, and communication style
- The difference between an interesting person and a useful next conversation
What one human cannot see
Attention is the scarce resource.
A person cannot inspect every profile, booth, proposal, and conversation. An agent can search a much larger field, then return a short list for human judgment.
The promise is not perfect recommendations. It is better coverage, more unusual adjacency, and more chances to notice what the room would otherwise lose.
Four lessons from the demo
What is worth carrying forward.
A useful introduction is a prediction.
Shared keywords may start the search. Context helps someone judge whether the conversation is likely to matter.
The gathering starts before the venue.
Profiles, questions, and possible connections can develop before people compete for attention in the room.
The platform's job is to protect the commons.
Identity, consent, scope, limits, and interruption are as important as what an agent can infer.
The value must come back to people.
The measure is not agent activity. It is a better human conversation and a clearer next action.
The hard part
Trust is the product.
Once personal agents can act around strangers, intelligence is only one design problem. The system also needs boundaries people can understand and organizers can enforce.
01Consent and participation
Who chose to participate? What exactly did they authorize? What happens to people who attend the physical event but do not join the digital layer? Participation cannot be assumed from a ticket purchase.
02Identity and scope
Every agent needs an accountable identity and a clear scope. The person should know what the agent can see, what it can do, and where its authority stops.
03Behavior and limits
Rate limits, moderation rules, and automatic pauses protect the shared environment. A useful system must be able to interrupt an agent, explain why, and give a human a path to repair.
04Privacy and visibility
What can an organizer inspect? Which messages are private? How long is information retained? Pattern filters can catch some personal information, but they are not a complete privacy strategy.
05Handoff and memory
If someone changes agents, what context should move with them? Memory can create continuity, but only when its ownership, limits, and deletion path are explicit.
06Human control
A recommendation should remain a recommendation. People need a clear way to inspect, reject, correct, and redirect what their agents do. The final social decision stays human.
The gathering system
Four conversations, not one.
We have been organizing for one visible interaction: a human speaking to other humans. Agent-enabled gatherings add three more paths. Each needs its own pacing, permission, and handoff.
The designer's job is to choreograph all four, then make their transitions legible to the people in the room.
See how Cognitor supports structured group thinking, explore designed conference introductions, or browse the Gathering Transformation Atlas.
The broadcast default is measurable
Nearly one in eight public agenda records looked stage-only.
United We Transform reviewed 23,624 public agenda records. In 12.1 percent of them, the public evidence showed a stage-only design: content moved outward, with no visible mechanism for participants to work with one another.
Read the State of Gatherings 2026 and see the evidence boundary.
Public agendas are incomplete records of an event. Missing public evidence is not proof that organizers did nothing elsewhere. The statistic describes what could be observed in the corpus.
From product to protocol
The demo pointed toward a shared layer.
Alistair's broader idea was a place where human-owned agents can interact around a shared domain even when they do not fully trust one another. The gathering is one application. The rules underneath may be useful far beyond conferences.
The AT Protocol came up as one possible substrate, not a decision. The important idea is separation: a shared communication layer should not trap every gathering inside one product.
The bigger picture
Gatherings need a new job description.
The point of a gathering is not to move content from a stage into rows of chairs. It is to create the conditions for people to notice one another, think together, and leave able to act.
The broadcast room
One voice. Everyone else receives.
Easy to schedule. Easy to recognize. Often poor at creating relationships or shared work.
- Attention points in one direction
- Participant knowledge stays hidden
- Chance encounters carry the networking load
The designed gathering
A host shapes the conditions for connection.
Greg described the role he wants: less broadcaster, more party host, with agents quietly helping him see which conversation the room needs next.
- People move between listening, making, and meeting
- Agents widen the field of possible connections
- Human hosts protect rhythm, safety, and meaning
The movement
Find the other misfits.
Alistair named the feeling plainly: this work can be lonely. Brandon answered with the reason United We Transform exists. More people are already pushing on this problem. We need to find one another faster.
Greg's internal map produced one introduction. Envoi tries to make that kind of contextual discovery possible across a room. The larger challenge is to redesign the room itself.
Sources and further reading
- Everything You Did to Make Your Conference Better Made It WorseMisha Glouberman's essay supplies the opening idea about microphones, chairs, and conference design.
- EnvoiOfficial product overview for the bring-your-own-agent event platform.
- BYOAI: Introducing EnvoiAlistair Croll's first-person account of the platform and its Startupfest debut.
- Envoi at StartupfestPublic participant walkthrough of profiles, talks, booths, messages, polls, and meeting recommendations.
- World ModelsResearch site by David Ha and Jürgen Schmidhuber, with background on internal models and selected representations.
- World Models paperResearch paper associated with the interactive site.
- AT Protocol overviewOfficial technical overview for the open protocol discussed as one possible substrate.
- Evidence for a collective intelligence factorWoolley and colleagues found that social sensitivity and more equal conversational turn-taking tracked collective performance in small laboratory groups. The study did not examine large conferences.
- Structural Holes and Good IdeasRonald Burt's research connects brokerage across disconnected groups with access to nonredundant ideas.
- Building connections at conferencesSensor research at two interdisciplinary conferences documented exploration, filtering, and persistent language homophily. It did not test agent matchmaking.
- State of Gatherings 2026United We Transform's public-agenda corpus and findings.
- Methodology and trust boundaryWhat the public evidence can and cannot support.
- About Greg LindsayBackground on Greg's work across cities, technology, and engineered serendipity.
- About Alistair CrollBackground on Alistair's work across technology and conferences.
- Primary conversationPrivate discussion among Brandon Klein, Alistair Croll, and Greg Lindsay on July 31, 2026. Product details are presented as a conceptual walkthrough unless linked to public product material.