Agentic Loops and the Headless Internet

ExplainPlan.ai began as a tool for turning scripts into animated presentation decks, but its creator realized the editor’s custom language still made people do too much work. The intended user could instead be an AI agent asked to produce a deck from a rough outline. To enable that workflow, the product added an MCP interface and an llms.txt file explaining its capabilities; a JavaScript-rendered site initially hid that guidance from agents, so it had to be made accessible without running JavaScript. Standard OAuth worked for web and terminal clients, while agents in virtual machines required a separate polling flow to receive access tokens. With those pieces in place, an agent could discover the product, authenticate, create and validate a deck, save it, and return a link.

The most consequential addition was a submit_feedback tool that let agents report failed tasks, bugs, confusing documentation, and feature ideas. After the agent could inspect rendered slides through a render_scene tool, it identified five visual defects that other validation checks missed, and requested batch rendering and clearer authentication behavior. Unlike people who may abandon a frustrating workflow without explanation, agents can describe exactly where they got stuck, giving product teams actionable roadmap input - though feedback still needs review and prioritization. The experience suggests products should treat agents as users, while preparing for bot-resistant access controls to impede legitimate delegated tasks and agent sessions to distort analytics and experiments. Teams may need to maintain machine-readable product guidance and separate agent traffic from human behavior; standards and agent capabilities are still evolving.