AI Engineer World's Fair 2026 — unofficial import demo

MCP v. CLI was about how agents talk to tools. That’s not settled (but we’re camp MCP… mostly). Almost nothing has settled how agents talk to each other - and that's where the next wave of value (and network effects and virality) lives. At Town we run a personal AI agent in production inside real people's inboxes, calendars, and Slack, and we've built agent-to-agent (A2A) on our platform: 1:1 A2A messaging, agents that carry a short bio of one another, HITL when sensitive data is shared or write actions are involved, and early tests around 1:N A2A. I’ll talk about the why, the opportunity, and the production architecture underneath. Audience takeaway: a concrete mental model for building multi-agent systems on top of the data and surfaces users already live in, plus our learnings on early failure modes to avoid.