Foundation Lab Method vs Durable Sessions: Which Should You Use?

What does the Emit Jane Luma Foundation Lab Method do?

The Foundation Lab Method is a strategic framework for building AI companies where product and research are not separate functions but one unified system. Developed from Luma AI's approach to visual and multimodal AI, it provides a complete playbook for how to couple model training with real customer usage data so that product improvements feed research and research improvements feed the product.

The method covers high-level decisions: Should you target filmmakers or marketers? Should you build a separate vision model or a unified multimodal model? Should you launch a consumer app or focus on enterprise deployments? Its core principles include thinking in professions rather than verticals, building the thinnest possible product stack on top of base model capability, capturing process data (how artifacts are made, not just the artifacts), and applying a logarithmic scaling test before major compute investments.

This framework is for founders, CEOs, research leads, and product leaders who are deciding the shape of an AI company — not for engineers solving a specific infrastructure problem.

What does the Christensen Durable Sessions Framework do?

The Durable Sessions Framework diagnoses and fixes a specific, concrete problem: AI chat and agent experiences that break under real-world conditions like network drops, multi-device usage, and concurrent agent activity. Developed by Mike Christensen of Ably, it identifies three foundational capabilities every production AI product needs — Resilient Delivery, Continuity Across Surfaces, and Live Control — and provides an architecture pattern to achieve all three.

The core architectural move is introducing a Durable Sessions layer between agents and clients. Instead of agents streaming directly to a client connection (the Single-Connection Trap), agents write events to a persistent, shared session. Clients subscribe to that session. This decoupling means streams survive disconnections, sessions follow users across devices, and clients can send steering or cancel signals without ambiguity.

This framework is for engineering leads, frontend architects, and infrastructure engineers building or auditing the real-time delivery layer of an AI product.

How do they compare?

These frameworks operate at entirely different layers of the AI product stack and are complementary, not competing.

The Foundation Lab Method asks: What should we build, for whom, and how should our company be structured so product and research compound each other? It is a strategic planning tool. You would use it in a board meeting, a strategy offsite, or a product-research alignment session.

The Durable Sessions Framework asks: Why does our AI chat break when users switch networks, and how do we fix the streaming architecture? It is a tactical engineering tool. You would use it in a sprint planning session, an architecture review, or a production incident postmortem.

The Foundation Lab Method would tell you to "build the thinnest possible product stack" — and the Durable Sessions Framework provides a specific architectural pattern for one critical layer of that thin stack. In fact, a Foundation Lab practitioner would recognize Durable Sessions as exactly the kind of infrastructure that avoids "fat stack" workarounds: it solves a delivery problem cleanly at the infrastructure layer rather than patching it with agent-side complexity.

On the dimension of complexity, the Foundation Lab Method is harder to apply because it requires organizational change — merging product and research, deploying Forward Deployed Creatives, restructuring data collection. The Durable Sessions Framework is more contained: it is a well-scoped infrastructure refactor with clear before-and-after testing criteria.

Which should you choose?

If you are deciding what AI company to build, how to structure your research-product loop, which professions to target, or whether to pursue a unified multimodal model, use the Foundation Lab Method. No amount of session infrastructure design will help you if your company is optimizing the wrong sub-problems or building separate product and research teams that pull in different directions.

If you have an AI product that works in demos but breaks in production — streams drop on mobile, users can't continue sessions across devices, the stop button doesn't work reliably, or your multi-agent orchestrator is a bottleneck — use the Durable Sessions Framework. Your strategic direction might be perfect, but users are churning because the experience is fragile.

If you are building a serious AI company from scratch, use both. The Foundation Lab Method sets the strategic direction and organizational structure. The Durable Sessions Framework ensures the product experience you ship is production-grade. They address different layers of the same challenge: building AI products that work in the real world.

The only scenario where you would explicitly choose one over the other is if your problem is purely strategic (you haven't started building yet, or you need to pivot your research-product relationship) versus purely tactical (your strategy is sound but your streaming architecture is broken). In most real-world AI companies, both problems exist simultaneously.

Can you use both frameworks together?

Yes, and this is the recommended approach for any AI company that is both training models and shipping products. The Foundation Lab Method's principle of "thin stack on top of base model capability" directly benefits from the Durable Sessions pattern — Durable Sessions is a thin, clean infrastructure layer that avoids the fat engineering harnesses the Foundation Lab Method warns against. Meanwhile, the Foundation Lab Method's emphasis on capturing process data from deployed agents requires a robust session infrastructure to capture those interactions reliably across devices and network conditions. The two frameworks reinforce each other.