Abstract

This disclosure documents the architecture of a production system that delivers a multi-step, AI-agent-driven content-production workflow to end users through two fundamentally different interaction surfaces at once: a real-time, stateful conversational interface (a chat bot, which inherently remembers where a conversation left off) and a fully stateless, asynchronous interface (individual links delivered by email, each a fresh, session-less HTTP request with no memory of any prior interaction). The system is architected so that a single, centrally persisted phase field is the sole source of truth for where a given user's in-flight workflow currently stands, with the chat interface consuming that phase directly and the email interface authenticating requests via tokens that every entry point independently revalidates against the live phase before acting, regardless of which channel a request arrives through. The workflow itself is decomposed into autonomous, multi-agent execution blocks (an AI-driven content-sourcing stage and an AI-driven drafting-and-revision stage) bounded by a small number of coarse-grained, re-entrant human-approval checkpoints, a structure chosen specifically because coarse, discrete decisions are what allow a checkpoint to be resolved identically regardless of which of the two channels the user happens to engage through. This disclosure describes the architecture, the specific engineering problems it solves (including a production incident and its fix), and an honest discussion of its relationship to the general, actively-developing pattern of human-in-the-loop AI agent checkpointing.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS