All your agents are going async
AI agents are evolving from simple chat windows to complex, asynchronous background processes, fundamentally breaking the synchronous HTTP transport model they were built upon. This technical deep dive explains why current solutions from major players like Anthropic and Cloudflare only partially address the problem. The author argues for a new 'durable transport' paradigm, proposing a session-based approach built on real-time messaging to truly enable persistent, flexible agent interactions.
The Lowdown
The world of AI agents is rapidly changing, moving from simple, synchronous chat-based interactions to complex, asynchronous background operations. This shift, while empowering agents to perform long-running tasks, schedules, and unattended workflows, exposes a fundamental flaw in their underlying communication mechanism: the HTTP request-response model. This traditional approach, designed for immediate responses, is ill-suited for agents that need to operate continuously, push unprompted updates, or survive connection disruptions.
- The Shift to Asynchronous Agents: Traditionally, LLMs operated like chatbots, streaming responses token-by-token over an open HTTP connection. However, agents are now integrating with crons, webhooks, and chat platforms like WhatsApp, becoming background workers that perform tasks while users are offline or busy.
- The HTTP Transport Mismatch: The HTTP request-response model cannot sustain long-lived agent work, handle agent-initiated pushes, manage caller changes (e.g., switching devices), or support multiple human collaborators in a single session.
- Industry Solutions: Companies like OpenAI (ChatGPT's scheduled tasks), Anthropic (Channels, Routines, Remote Control), and Cursor (background agents) are attempting to solve this. OpenClaw notably demonstrated agents living within WhatsApp, effectively separating agent work lifetime from human connection lifetime.
- Partial Solutions: Durable State vs. Durable Transport: Many current approaches, including those from Anthropic and Cloudflare (with their Agents platform and Sessions API), primarily focus on providing durable state for agents. They store conversation history and agent progress but often rely on polling or basic HTTP for actual data transport, failing to solve the durable transport problem.
- The Missing Piece: Durable Transport: The author, affiliated with Ably, argues that a truly robust asynchronous agent system requires both durable state and a durable transport. This transport would allow agents and humans to connect and disconnect seamlessly, survive network issues, and enable server-initiated pushes, effectively treating an AI 'session' as a first-class, persistent primitive.
In essence, the move towards truly autonomous and persistent AI agents demands a complete rethinking of their communication infrastructure. Until a durable, real-time transport mechanism that inherently supports continuous, multi-party, and fault-tolerant sessions is adopted, the full potential of asynchronous agents will remain hindered by the limitations of a synchronous, short-lived communication paradigm.