Beyond HTTP: The Need for Durable Transport in Async AI Agents

AI agents are transitioning from synchronous chat tools to asynchronous background workers, making traditional HTTP-based transport obsolete. While some platforms are improving state storage, they still struggle with the transport mismatch that prevents reliable, multi-device, and unprompted communication. A durable, session-based messaging transport is essential to support the long-running and decoupled nature of modern AI agents.

Key Points

• AI agents are moving away from synchronous chat toward asynchronous background routines and scheduled workflows.
• The traditional HTTP request-response model is inadequate for async agents because it cannot survive connection drops, page refreshes, or device hand-offs.
• The infrastructure challenge is split into two parts: 'durable state' (storing agent data) and 'durable transport' (delivering the data reliably).
• Current solutions from major providers focus on state but still rely on inefficient polling or fragile HTTP connections for transport.
• A session-based, bi-directional messaging platform is necessary to provide the durable transport required for modern, multi-device agentic applications.

Sentiment

The community is skeptical of the article's proposed solution while acknowledging the underlying problem. Most commenters agree that async agents are an important trend, but they view the transport layer as either already solved or as the wrong layer to optimize. The strongest pushback targets the article as a product pitch and argues that context management, session portability, and agent governance are more pressing unsolved problems than message delivery.

In Agreement

• The shift from synchronous chat to async background agents is real and already happening across major platforms like Claude, Codex, and Devin
• HTTP request-response with SSE streaming breaks when agents outlive connections, and page refreshes during streaming cause real UX problems
• Multi-device and multi-user agent sessions are genuinely hard to build on top of traditional HTTP patterns
• The security and governance surface expands significantly when agents run asynchronously without human supervision
• A durable messaging layer that supports disconnect/reconnect would improve the agent development experience

Opposed

• The article is fundamentally a sales pitch for Ably's pub/sub product dressed up as a technical blog post
• Websockets plus server-side state persistence already solve the transport problem without needing a new abstraction — this is how Codex and other production systems work
• The real problem isn't transport but context management — agents need the ability to dynamically compose and prune their context window rather than just concatenating messages
• Existing tools like Temporal, Cloudflare Durable Objects, and Google's ADK already handle durable async execution
• Session state portability between different agent tools is the actual blocker, not the messaging layer
• Agents are getting faster which keeps synchronous interaction relevant, and async has limited utility until agents become more reliable on long-horizon tasks
• Simple solutions like SSH plus tmux, email, or polling with request IDs handle async agent communication without new infrastructure