MCP Tasks: Why Agents Aren't Supporting Them

Cornelia Davis, a Principal Technologist at Temporal, recently addressed the perplexing question of why MCP tasks are not yet widely supported by agents. Speaking at the AI Engineer World's Fair, Davis highlighted that the complexity and experimental nature of the MCP Tasks specification, first released in November, are the primary reasons behind this hesitancy.

Understanding MCP Tasks Complexity

Davis explained that the initial version of the MCP Tasks protocol (V1) was marked as experimental, which often leads developers to adopt a wait-and-see approach. She elaborated on the inherent complexities, particularly those related to stateful protocols and the challenges of handling long-running tasks. These include dealing with infrastructure issues like network disruptions, server crashes, and the need for tasks to remain durable even when clients or servers go offline.

A key challenge in V1, as Davis pointed out, was the 'task_list' endpoint, which was stateful and lacked filtering capabilities. This meant that with a large number of agents and tasks, clients would have to sift through millions of tasks to find the one they needed to interact with. Furthermore, the 'task_result' endpoint, used for eliciting responses, involved a complex protocol with long-running connections that proved difficult to manage, especially in cases of connection failures.

MCP Tasks V2 and Future Improvements

Davis then introduced MCP Tasks V2, which is slated for release in July and promises significant improvements. A major change is the move towards a stateless core, which is a significant simplification for distributed systems. The protocol will also feature extensions for server-rendered UIs via MCP Apps and long-running workflows through the Tasks extension. Additionally, authorization will be more closely aligned with OAuth and OpenID Connect, and a formal deprecation policy will be introduced to ensure smoother protocol evolution.

During her presentation, Davis showcased a live demo illustrating a purchase order processing use case. This demo highlighted how MCP tasks enable long-running, asynchronous operations, such as validating invoices against an ERP system, requesting human approval, and reconciling data. She demonstrated the workflow, including the ability to signal into a long-running process, showcasing the power of these asynchronous task capabilities.

The Importance of Durability and State Management

A crucial aspect emphasized throughout the talk was the concept of durability. Davis quoted the spec, stating that "Tasks are durable state machines that carry information about the underlying execution state of the request they wrap." This means tasks must survive infrastructure failures and allow interaction once systems are back online. She stressed that clients SHOULD persist task IDs to durable storage to ensure polling can resume after crashes or restarts, noting that while the spec recommends this, it's not yet a strict requirement.

Finally, Davis touched upon the scaling challenges, particularly the inefficiency of millions of clients polling for task updates simultaneously. She hinted at upcoming improvements in the notification part of the tasks protocol, which would allow for a more efficient, event-driven approach rather than constant polling. She also mentioned that the experimental code for this is already open source and that Temporal is working on providing both client and server implementations for the protocol layer, aiming to make it easier for developers to adopt.