Durable Sessions vs Agentic Web Publishing: Which Framework?

What does the Christensen Durable Sessions AI UX Framework do?

Mike Christensen's framework diagnoses why AI chat and agent products feel fragile in production and provides a concrete architectural fix. The core insight is that most AI products use direct HTTP streaming (typically SSE via tools like the Vercel AI SDK) where the response stream is coupled to a single client connection. When that connection drops — a mobile user switches networks, a laptop lid closes, a tab refreshes — the stream is gone forever.

The framework identifies three foundational capabilities that separate a demo from a production-quality AI experience: Resilient Delivery (streams survive disconnections), Continuity Across Surfaces (sessions follow users across tabs and devices), and Live Control (users can steer, interrupt, or cancel agents mid-generation). It then prescribes a specific architectural pattern — Durable Sessions — where a persistent, shared, independently addressable channel sits between agents and clients. Agents write events to the session. Clients subscribe to the session. Neither holds a direct reference to the other.

This framework is especially powerful for multi-agent architectures, where it eliminates the Orchestrator Dual-Purpose Problem by letting every sub-agent publish directly to the session channel instead of routing all updates through a central orchestrator.

What does Rachel Lee's Agentic Web Publishing Framework do?

Rachel Lee's framework solves a completely different problem: how to make existing web content and functionality accessible to AI agents. Her premise is that your audience is already working inside agents like Claude Desktop, Copilot, and in-browser AI assistants. If your content only exists as traditional HTML pages, those agents cannot meaningfully interact with it.

The framework introduces the One Server, Three Clients principle: a single backend should serve humans in browsers, humans operating agents, and agents browsing autonomously. You achieve this by auditing your site's navigation and mapping each action (list, search, filter, view detail) to an MCP tool exposed via an HTTP transport endpoint. For tools that benefit from rich interaction, you return MCP Apps — self-contained HTML/CSS/JS bundles rendered inline inside the agent harness — rather than plain JSON. For in-browser agents, you add Web MCP attributes to your existing HTML pages so agents can call your site's functions directly instead of guessing from screenshots.

A key design principle is that chat-only interfaces (what Lee calls the "starfish design") put all discovery burden on the user. MCP Apps and structured tool outputs provide visual cues, guided navigation, and interactive experiences that go far beyond walls of text.

How do they compare?

These two frameworks operate at entirely different layers of the stack and solve different problems:

Durable Sessions is an infrastructure framework. It answers: "How do I make my AI product's real-time streaming reliable, multi-device, and controllable?" It requires significant architectural work — replacing SSE, introducing a pub/sub session layer, rewiring agent output and client subscriptions. The payoff is a fundamentally more resilient and capable product.

Agentic Web Publishing is a distribution framework. It answers: "How do I make my existing web content reachable by AI agents?" It is incremental and additive — you can ship one MCP tool without changing your existing site. The payoff is reaching users who now spend their time inside agent harnesses rather than browsing the open web.

They do not compete. A team building an AI coding assistant with multi-agent streaming would use Durable Sessions for the infrastructure. That same team could also use Agentic Web Publishing to expose their documentation site to agents. In fact, a mature AI product will likely need both: robust delivery infrastructure (Durable Sessions) and broad agent-accessible distribution (Agentic Web Publishing).

Where they overlap is in philosophy: both reject the naive default. Christensen rejects the assumption that a direct SSE pipe is sufficient. Lee rejects the assumption that a chatbox is a complete interface. Both push toward richer, more resilient AI experiences.

Which should you choose?

Choose Durable Sessions if you are building or operating an AI product with real-time streaming responses and you are experiencing any of these symptoms: responses lost on mobile network switches, no multi-device session continuity, a stop button that sometimes kills the stream instead of stopping generation, or an orchestrator drowning in relay logic. This is your framework. It is higher effort but addresses problems that will only get worse as your product scales.

Choose Agentic Web Publishing if you have a website, documentation site, content archive, or SaaS dashboard and you want to open it up to AI agents. This is especially urgent if your analytics show declining direct traffic and you suspect users are accessing information through AI assistants instead. The incremental approach (start with one HTTP MCP tool, then add MCP Apps) makes it low-risk to begin.

Choose both if you are building a full-stack AI product that includes both a real-time agent experience (needing resilient streaming) and web-published content or tools (needing agent accessibility). Start with whichever addresses your most acute pain point, then layer in the other.

If you are forced to pick one starting point: the Durable Sessions framework addresses problems that actively break user experiences today, while Agentic Web Publishing addresses a distribution opportunity that is growing but not yet urgent for every team. Fix what is broken first, then expand where you are found.