Durable Sessions AI UX vs Ng Deep Learning Execution

What does the Christensen Durable Sessions AI UX Framework do?

The Durable Sessions AI UX Framework, introduced by Mike Christensen of Ably, diagnoses why AI chat and agent-driven product experiences break under real-world conditions — and provides a concrete architectural pattern to fix them. The core insight is that most AI products stream responses over a direct HTTP connection (typically SSE via tools like the Vercel AI SDK), which couples the health of the entire response to a single client connection. If that connection drops — a mobile user switches networks, a laptop lid closes, or a tab is backgrounded — the stream is gone.

The framework identifies three foundational capabilities that separate a fragile demo from a production-quality AI product: 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). The solution is a Durable Sessions layer — a persistent, shared pub/sub channel between agents and clients where agents publish events and clients subscribe. Neither holds a direct pipe to the other. This architectural inversion also solves the SSE Resume-Cancel Conflict (where closing a connection is ambiguous between a disconnect and a user cancel) and the Orchestrator Dual-Purpose Problem in multi-agent systems.

This framework is purely about delivery infrastructure. It assumes your model and agent logic already work. Its value is in making that working logic actually reach users reliably.

What does the Ng Deep Learning Project Execution Skill do?

Andrew Ng's Deep Learning Project Execution Skill, drawn from Stanford's CS230 course, provides a systematic methodology for designing, diagnosing, and accelerating AI and deep learning projects. It addresses the most common reason ML projects stall: teams making random decisions about what to work on next — collecting more data one week, buying GPUs the next — without running diagnostics first.

The framework walks practitioners through classifying their application by data type and abstraction layer (prompting vs. deep learning vs. traditional ML), honestly assessing their data situation, building quick-and-dirty prototypes in sandboxed environments, and then — critically — running diagnostics before choosing interventions. It provides a prioritized intervention hierarchy: fix data quality first, tune hyperparameters second, adjust architecture third, fine-tune a pre-trained model fourth, and only buy more compute after exhausting the above.

Ng's skill also addresses cost curve awareness — knowing when to drop from expensive LLM API calls to fine-tuned smaller models — and the principle that fast iteration in a sandbox is itself a safety mechanism. This framework is about getting the model and the project right. It does not address how the model's output reaches users.

How do they compare?

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

Durable Sessions lives at the delivery and infrastructure layer. It answers: "My model works, but users experience broken streams, can't resume after disconnections, can't see responses on a second device, and can't stop a generation cleanly." It is an architectural pattern for real-time AI product engineering.

Ng's Project Execution Skill lives at the model development and project management layer. It answers: "I have an AI project idea (or a stuck project) and I need to systematically figure out what to build, what data I need, and where to invest effort to reach target performance." It is a diagnostic and execution methodology for ML/DL work.

The overlap is near zero. A team struggling with model accuracy gains nothing from Durable Sessions. A team whose model is excellent but whose users lose responses on mobile gains nothing from hyperparameter tuning diagnostics.

The one area of indirect overlap is that both frameworks share a bias toward disciplined diagnosis over random action. Christensen's framework audits streaming architecture against specific failure modes before prescribing changes; Ng's framework runs error analysis before prescribing interventions. Both oppose the "try random things" approach.

Which should you choose?

Choose the Durable Sessions AI UX Framework if you already have a working AI agent or LLM-powered product and your problems are about delivery: dropped streams, no multi-device support, no stop button, SSE limitations, or multi-agent progress visibility. This is Christensen's framework and it is clearly better for real-time infrastructure and AI UX architecture problems.

Choose the Ng Deep Learning Project Execution Skill if you are starting an AI project, your model isn't performing, you're stuck choosing between more data vs. more compute, or you need to decide whether to use LLM prompting or train a custom model. Ng's framework is clearly better for ML project scoping, diagnosis, and execution.

Use both if you are building an AI product end-to-end. Use Ng's skill to get the model right, then use Durable Sessions to make sure users actually receive the model's output reliably across real-world conditions. Most serious AI product teams will need both skill sets — the model layer and the delivery layer — and conflating the two is itself a common mistake.

If you are forced to pick one to learn first, start with whichever matches your current bottleneck. If your model doesn't work, no amount of infrastructure will save you. If your model works but users complain about broken experiences, no amount of hyperparameter tuning will fix it.