Keynote at ETH Juniors Momentum

The Agentic Systems Lab recently delivered a keynote on agentic AI at ETH Juniors' Momentum event in Zurich, exploring what agentic AI is, where it is going, and how to build it and with it. The Momentum event gathered Swiss investors, founders, and executives to discuss the latest developments in technology.

The keynote was delivered by Kevin O'Sullivan, Robert Jakob, and Connor Larson, and opened with a deliberately provocative claim for an agentic AI lab: most businesses need fewer AI agents than they think, and each one should do far more. As context windows have crossed the million-token threshold, a single agent running many turns and holding shared context across tasks often outperforms a swarm of narrow, task-specific agents. It costs less, is simpler to build and observe, and, because it keeps the target schema in context, produces structured output that fits existing systems on the first pass.

The talk also reframed the widely cited statistic that most enterprise generative AI pilots fail. With agentic engineering, technical and non-technical users alike can prototype quickly, so the winning strategy is to try many things and keep only the winners.

The speakers also argued that the harder problem is productionizing, which makes the case for building eval-first. Rather than relying on vibes, the approach is to instrument everything, measure what actually works, and let evaluations decide what to harden and ship.

The keynote also closed on where to start. Organizations that have already rolled out tools like ChatGPT, Claude, or Copilot can watch where their people naturally reach for AI during the day, and those recurring, high-value workflows form the shortlist for what to build first.

The keynote brought together Swiss investors, founders, and executives at ETH Juniors' Momentum event to discuss the latest developments in technology and the trajectory of agentic AI.