What It Really Takes to Build AI Agents

Our founder Chris Van Dyke recently joined Michael Domanic, Head of AI at Section, for a live session as part of Section's AI: ROI Conference: What It Really Takes to Build AI Agents.

The premise of the conversation is one we live every day at SplittingAtom: most companies experimenting with AI agents can get to something workable. Far fewer know how to turn that momentum into something durable, governed, and genuinely useful. The distance between a promising demo and an agent you actually rely on is not a bigger model — it is memory, identity, oversight, and the unglamorous infrastructure around them.

The session covered the foundational elements agents need before they can be trusted with real work, how to design agents that deliver measurable output rather than impressive transcripts, and what it takes to scale a successful experiment across an organization — including the governance questions of autonomy, risk, and accountability that determine whether agents create real business value or just activity.

If you are wrestling with the same gap between prototype and production, the full recording is available on Section's site. And if you want to talk about what it takes to give agents memory and a place to work, get in touch.