CJ Brewer, Co-founder at Swirls AI

The talk

Could this just be a markdown file? Declarative agents for real operations

How to take an agent from “roughly follow these instructions” to a declarative operational system with typed contracts, scoped tools, and human approval points.

Most agent workflows can begin as a Markdown file: instructions, examples, a few tools, and a desired outcome. That is useful and often the right place to start. But a business cannot run on “roughly follow these instructions.” Once an agent touches customer data, sends emails, updates systems of record, or triggers work for a team, operational builders need to answer harder questions. What inputs are allowed? What outputs are guaranteed? Which actions can the agent take? What happens when something fails? Where does a human approve?

This talk makes the case that the path from an agent prompt to an operational system is declarative. Make the agent’s inputs, outputs, tools, permissions, triggers, workflows, and approval points explicit. Then the system becomes inspectable, versionable, testable, and deployable. In engineering terms, we are bringing a familiar idea into agentic development: constrain the variable reasoning layer inside a defined contract.

We’ll walk through how an inbox-to-CRM agent evolves from a Markdown instruction into a production-ready workflow with typed inputs and outputs, deterministic tool calls, scoped authority, durable execution, and human review. The thesis: businesses will increasingly run day-to-day through agents. The teams that succeed will not treat those agents as prompts. They will treat them as declarative operational systems.