The prose is the source. Across current service-automation platforms, the source of truth for agent behavior is moving away from hand-built decision trees and toward written procedures: Decagon compiles operating procedures written in plain language into executable agent logic, Intercom’s Fin follows written procedures, Sierra drafts agents from conversation, and Parloa works from briefing documents. What is still largely missing is the editing environment around that shift — a place where a subject-matter team can see what their prose compiles to, test it against realistic conversations, and understand what a wording change does before it reaches customers. This studio sketches that environment.
The task model as control point. The middle panel follows a second line of work: studies on generative interfaces (the Jelly system, 2025; related work at Google Research) argue for exposing an editable task model between intent and interface, so that the model — not the prompt — becomes the thing people steer. Here the extracted model writes back into the prose: change the refund limit on the card and the sentence changes with it, and the next test run behaves differently.
Scope, honestly. The compiler is a deterministic pattern parser over sentence structure — conditions, actions, verification steps, escalations, numeric limits and never-rules. It is a design probe for the editor, test and diff interactions, not a language-understanding system; in production the same surfaces would sit on top of a real procedure compiler.
References: decagon.ai/product/aop (agent operating procedures) · fin.ai (procedures, outcome pricing) · sierra.ai (agent drafting, supervision) · parloa.com (briefing-based agents) · Jelly — dl.acm.org/doi/10.1145/3706598.3713285 · Ink & Switch, “Malleable software” (2025). Built by Dr. Raoul Pilcicki as part of a five-part series of working interfaces for agentic service delivery.