Durable Sessions vs Best Agents Four-Pattern: Which Framework?
// TL;DR
Use Swanepoel's Four-Pattern Framework if you are designing or improving an AI agent's interaction quality, trust, and personalization. Use Christensen's Durable Sessions Framework if your AI product already works well in demos but breaks under real-world network conditions, multi-device usage, or multi-agent streaming. Most teams should start with Swanepoel's patterns to get the agent experience right, then layer on Durable Sessions to make delivery resilient. They solve different layers of the stack and are complementary, not competing.
// HOW DO THEY COMPARE?
| Dimension | Christensen Durable Sessions AI UX Framework | Swanepoel's Best Agents Four-Pattern Framework |
|---|---|---|
| Best for | Fixing broken streaming, disconnections, and multi-device AI delivery | Improving agent output quality, user trust, personalization, and safety |
| Layer of the stack | Infrastructure and transport layer (between agent and client) | Application and interaction design layer (agent behavior and UX) |
| Complexity to implement | High — requires architectural changes to streaming, transport, and session management | Medium — requires prompt engineering, UI changes, and workflow redesign but no infrastructure overhaul |
| Time to apply | Weeks to months — involves replacing SSE, adding pub/sub, redesigning connection model | Days to weeks — Focus Modes and Transparent Execution can be shipped incrementally |
| Prerequisites | Existing AI product with streaming architecture; understanding of SSE, WebSockets, pub/sub | Any AI agent at any stage; no specific infrastructure requirements |
| Output type | Architectural redesign plan with session layer, transport changes, and validation tests | Agent design audit with scored patterns and concrete improvement roadmap |
| Primary problem solved | Streams die on disconnect; no multi-device sync; stop button is ambiguous | Low user trust; generic outputs; irreversible mistakes; unfocused agent behavior |
| Multi-agent support | Excellent — directly solves orchestrator relay bottleneck for multi-agent progress streaming | Not addressed — framework is agent-agnostic but does not cover multi-agent coordination |
| Creator background | Mike Christensen, Ably — real-time infrastructure and messaging platform specialist | Mardu Swanepoel, Flinn AI — AI agent design and product experience specialist |
| When to skip | Skip if your product is not yet streaming or if agent quality is the bottleneck, not delivery | Skip if your agent design is solid but streams keep breaking under real-world network conditions |
What does the Christensen Durable Sessions AI UX Framework do?
Christensen's Durable Sessions Framework diagnoses and fixes the infrastructure layer between your AI agents and your users. It targets a specific failure: the Single-Connection Trap, where your streaming architecture (typically SSE) couples the health of the AI response to a single client connection. When that connection drops — a mobile network switch, a tab refresh, a flaky Wi-Fi signal — the stream dies and the user loses their response.
The framework introduces a Durable Sessions layer, a persistent shared resource that sits between agents and clients. Agents write events to the session; clients subscribe to the session. Neither holds a direct pipe to the other. This architectural inversion delivers three foundational capabilities: 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 framework also solves the SSE Resume-Cancel Conflict — the ambiguity where closing an SSE connection could mean either "I disconnected, let me resume" or "I pressed stop, cancel generation" — by requiring bidirectional transport like WebSockets. For multi-agent architectures, it eliminates the Orchestrator Dual-Purpose Problem by letting sub-agents write directly to the session rather than routing all progress updates through a bottlenecked orchestrator.
This is an infrastructure framework. It does not address whether your agent produces good outputs — it ensures those outputs reach users reliably, across devices, with full user control.
What does Swanepoel's Best Agents Four-Pattern Framework do?
Swanepoel's framework identifies the four structural patterns shared by the world's best AI agents and provides a systematic method to audit and improve any agent against them. The four patterns are:
1. Focus Modes — Constrain the agent into named modes (e.g., Research Mode, Debug Mode) that narrow the action space, optimize prompts, and set clear user expectations. A do-anything agent does nothing well.
2. Transparent Execution — Surface the agent's process, not just its results. Show a live to-do list, tool calls, assumptions, and uncertainties. This shifts the user from blind delegation to active collaboration and builds trust.
3. Personalization — Equip the agent with the user's own playbooks, memories, and connected knowledge so it produces outputs that look like the user did them, not a generic AI.
4. Reversibility — Make every consequential agent action undoable at multiple levels of granularity. When users know the downside is bounded, they authorize bolder, higher-value tasks.
The framework is practical and incremental. You score your agent on each pattern (absent, partial, present), identify the highest-leverage gap, and improve pattern by pattern. It works at any stage of agent development and requires no specific infrastructure.
How do they compare?
These frameworks operate on completely different layers of the AI product stack and solve fundamentally different problems.
Christensen's Durable Sessions Framework is an infrastructure and transport solution. It answers: "How do I get the agent's output to the user reliably, across devices, with live control?" It is deeply technical, requires architectural changes, and is most relevant to teams whose agent works well in a demo but breaks in production under real-world network conditions.
Swanepoel's Four-Pattern Framework is an interaction design and agent behavior solution. It answers: "How do I make the agent's output trustworthy, personalized, focused, and safe?" It is applicable to any team building or improving an agent, regardless of infrastructure maturity.
The key distinction: Durable Sessions assumes your agent is producing good output and fixes the delivery pipe. The Four-Pattern Framework assumes delivery works and fixes the agent's behavior and UX. Neither replaces the other.
For multi-agent architectures, Christensen's framework is clearly stronger — it directly addresses the orchestrator relay bottleneck and multiplexed progress streaming. Swanepoel's framework does not address multi-agent coordination at all.
For improving trust, output quality, and user engagement, Swanepoel's framework is clearly stronger. Transparent Execution and Personalization are high-impact UX improvements that Durable Sessions does not touch.
Which should you choose?
Start with Swanepoel's Four-Pattern Framework. Most AI agents fail because of poor output quality, lack of user trust, generic results, and irreversible mistakes — not because of streaming infrastructure. The Four-Pattern Framework addresses these problems directly, ships incrementally, and requires no infrastructure overhaul. It is the higher-leverage starting point for the majority of teams.
Add Christensen's Durable Sessions Framework when your agent is good but delivery is fragile. If users report lost responses on mobile, if your multi-device experience is broken, if your stop button is unreliable, or if your multi-agent orchestrator has become a tangled relay — Durable Sessions is the precise fix. It is a necessary investment for production-grade AI products but should not be your first priority if the agent itself is underperforming.
For production AI products at scale, you need both. The best AI products combine excellent agent behavior (Four-Pattern) with resilient, multi-surface delivery (Durable Sessions). Apply Swanepoel's patterns to get the agent right, then apply Christensen's architecture to make sure users reliably experience that quality.
If you are building a multi-agent system with real-time progress streaming, prioritize Durable Sessions earlier — the Orchestrator Dual-Purpose Problem will create compounding technical debt if not addressed at the architecture level.
// FREQUENTLY ASKED QUESTIONS
Can I use Durable Sessions and the Four-Pattern Framework together?
Yes, and you should for production AI products. They solve different layers: Swanepoel's framework improves what the agent does and how users interact with it, while Christensen's framework ensures the agent's output reaches users reliably across devices and network conditions. They are complementary, not competing.
Which framework should I use if my AI agent gives mediocre answers?
Use Swanepoel's Four-Pattern Framework. Mediocre answers are an agent behavior problem — likely caused by missing Focus Modes (agent is too unfocused), missing Personalization (outputs are generic), or missing Transparent Execution (you can't diagnose where quality breaks down). Durable Sessions won't help with output quality.
Which framework fixes my AI chat losing responses when users switch networks?
Christensen's Durable Sessions Framework directly solves this. Lost responses on network switch are the defining symptom of the Single-Connection Trap. The fix is a persistent session layer that buffers events and lets clients reconnect and resume exactly where they left off, without any agent-side replay logic.
Do I need to replace SSE to implement Durable Sessions?
If you need Live Control (stop buttons, steering messages mid-generation), yes — SSE is one-way, creating the Resume-Cancel Conflict. You need bidirectional transport like WebSockets. If you only need Resilient Delivery and Cross-Surface Continuity, a pub/sub layer over SSE can partially work, but WebSockets are recommended for the full capability set.
How long does it take to implement the Four-Pattern Framework?
Days to weeks for meaningful improvement. Focus Modes can be added by splitting system prompts and tool sets per mode. Transparent Execution requires UI work to surface agent progress. Personalization needs playbook or memory integration. Reversibility requires undo mechanisms. Each pattern can be shipped independently and incrementally.
Is Swanepoel's framework only for chatbot-style agents?
No. The four patterns apply to any AI agent — coding assistants, research agents, customer support bots, document processors, autonomous workflow agents. Focus Modes, Transparent Execution, Personalization, and Reversibility are structural patterns that improve any agent where users interact with or depend on agent output.
What is the biggest mistake teams make when building AI agent products?
Both frameworks identify a common root cause: focusing on model quality while ignoring the experience layer. Christensen warns that the gap between demo and production is almost entirely infrastructure. Swanepoel warns that speed to outcome without speed to understanding produces fast but useless results. Both argue the experience layer is underinvested.
Which framework is better for multi-agent architectures?
Christensen's Durable Sessions Framework is clearly better for multi-agent architectures. It directly solves the Orchestrator Dual-Purpose Problem where the orchestrator becomes a bottleneck relaying sub-agent updates. With Durable Sessions, every sub-agent writes directly to the shared session. Swanepoel's framework does not address multi-agent coordination.