The Multimodal Communication Protocol (MCP), a key standard for enabling Large Language Models (LLMs) to interact with external tools and their context, is undergoing a significant transformation. Historically, MCP's primary function was to provide a standardized mechanism for an LLM (the model) to connect with and invoke various tools (e.g., APIs, databases, external services) and to manage the context of these interactions. However, recent developments outlined in the MCP maintainers' new roadmap suggest a much broader ambition, extending its role to that of a foundational agentic infrastructure.
One of the most notable shifts is the recognition that modern agentic workloads no longer fit the traditional request-response model. Agents are expected to perform long-running tasks that require continuous interaction and feedback. To address this, MCP is introducing primitives for:
Streamed Results: Instead of waiting for a single, final response, agents will be able to receive streamed results as their tasks progress. This is crucial for agents that need to provide real-time updates or adapt their behavior based on intermediate outcomes.
Steering and Events: The protocol will enable dynamic steering of agents during execution, allowing external systems or other agents to guide their actions. Concurrently, servers will gain the ability to push events back to agents, transitioning from a purely pull-based interaction (where clients constantly query for status) to a more efficient, event-driven model.
Beyond execution, the roadmap also highlights advancements in agent identity and delegation. Current authentication mechanisms, like OAuth, inherently assume a human user in a browser approving actions. This model breaks down when the caller is a cloud agent acting on behalf of a user, or when an agent needs to delegate limited authority to a subagent. MCP aims to standardize identity and delegation specifically for these agent-to-agent and agent-to-service interactions, which is a critical step towards building secure and robust multi-agent systems.
Another significant improvement is progressive tool discovery. In existing setups, if an MCP server exposes many tools (e.g., 100+), the model has to incur the context cost of loading the descriptions of all these tools, even if only a few are relevant to the immediate task. This bloats the context window and can degrade tool selection performance. The proposed solution involves allowing the server to expose a smaller, initial surface of tools and then progressively reveal more tools as the agent's intent becomes clearer. This dynamic discovery mechanism will lead to more efficient context management and improved tool utilization.
In essence, MCP is evolving from a simple protocol for calling tools into a comprehensive execution substrate for long-running agents. This means it's becoming less about what tools an LLM can call and more about how agents operate, interact, and manage their lifecycle and security within a broader ecosystem. For senior software engineers, this implies that designing agentic systems will increasingly involve leveraging these new MCP capabilities to build more sophisticated, resilient, and autonomous agents, where the distinction between the agent's core logic and its interaction framework becomes tightly integrated. This evolution changes how we think about agent architectures, from basic function calling to complex, distributed, and adaptive agentic workflows.