Why agents can't yet handle durable async workflows—and what's changing

MCP tasks solve a real problem: long-running agent workflows that need to survive infrastructure failures, network blips, and human delays. Cornelia Davis demonstrates this with a purchase order use case, showing how tasks move through lifecycle states like working, input required, and completed. The challenge is that durability requires both server-side and sophisticated client-side implementations.

The V1 specification introduced stateful protocols and complex long-running connections for handling human-in-the-loop inputs, which made client adoption impractical. Most teams building agents have wisely avoided implementing it. The upcoming V2 specification, announced in May 2025, moves the protocol to a stateless core, removes the problematic task list endpoint, and replaces long-running connections with a

simple signal-based update mechanism. This dramatically simplifies client implementation. However, even V2 has unresolved scalability concerns: polling millions of tasks individually doesn't work. A notifications-based protocol is proposed but not yet specified. The speaker built her own client implementation using Temporal workflows to handle the complexity, highlighting that practical adoption requires significant engineering investment. For

engineering leaders, the key tension is between the clear need for async long-running agent tasks and the immature infrastructure supporting them. The speaker expects fast MCP to offer simpler abstractions soon, but production-grade support is still months away. Teams should watch this space but not bet on it today. The fundamental insight is that agentic