The Agent Is Not the Product

Every production agent system is a mix of agent and workflow. The agent does the reasoning; the workflow holds the state, the sequencing, the policy, and what happens when things go sideways. Almost all the money and attention goes to the first half.

This session argues the second half is the real product, and whoever builds it properly becomes as foundational to the AI stack as cloud providers became to the web. I'll set out the six things a control plane has to do: identity and access, spend governance, step-level policy enforcement, observability, audit and replay, and human oversight designed in rather than bolted on — showing what it looks like when each is missing in systems that have already failed publicly.

You'll leave with a rubric for those six, and scores for major agent platforms drawn from public documentation, so you can run the same assessment yourself. n8n is in the same table, under the same rules.