The WeAreDevelopers World Congress came to North America for the first time on September 23 to 25, 2026, at the San José McEnery Convention Center, with more than ten thousand attendees and Docker as presenting partner. The series has been held in Europe for years; this was its North American debut. Alexy Khrabrov was there for the final day.
His read on the hall, recorded on the floor: it is all AI, all agents, all AI thought leadership, and how you cope with that much change. Plus the startups in uniforms selling their tooling, which he found instructive, and a great many good people in the audience.
Two of those people he interviewed, and both were arguing about the same layer from different sides.
Kranti Parisa, co-founder of LaserData, made the case for what he and Alexy call AI Infra 2.0. The argument is that agentic infrastructure is still being built on legacy stacks, and that every hop across a JVM boundary forces serialization, so Amdahl's law caps whatever parallelism can recover — efforts to accelerate JVM systems reach single-digit speedups where Rust systems reach ten to a hundred times. LaserData's streaming engine, built on Apache Iggy, is Rust, with millisecond P99 response times and three to four times the throughput. The part worth arguing with is his claim about what actually costs money in production: not inference, but context propagation. An agent working from stale context wastes compute, multiplies subsystem calls, and pollutes observability data, and that is a latency problem before it is a token problem.
Vlad Luzin, co-founder and CTO of BAND, is working a layer up: letting agents talk to each other. Connecting agents, he points out, is a distributed system of microservices, and most of the work is plumbing. BAND supplies the primitives — a registry in which agents discover each other and work as a group but cannot reach outside it without consent, an identity for each agent, and token and credential propagation so one person's agent acts in that person's scope. The registry is a deterministic data model rather than a compile-time decision, and it belongs to a human: every agent, coding assistants included, has an identity and an owner, and the owner decides who joins.
The recordings are below, and the photographs from the day are on f.photo.