Exploring MCP Primitive Usage Beyond Designed Use Cases: Tool Calling in Python

Unlocking the Power of Agent-to-Agent Communication with MCP: A Game-Changer for Autonomous Agents and Observability

1. The speaker is discussing MCP (Microagent Communication Protocol) and how it can be used for agent-to-agent communication.
2. MCP was designed primarily for use in autonomous agents or code that interacts with other code, not for UI coding agents such as web browsers or desktop applications.
3. The speaker is a creator of Pyantic, a data validation library for Python, and is also involved with the MCP Python SDK.
4. Pantic has built two additional tools: Pantic AI, an agent framework for Python, and Pantic Logfire, an observability platform for commercial use.
5. The speaker clarifies that neither MCP nor Pyantic are all-encompassing tools, but they can handle a significant amount of tasks in the right context.
6. MCP consists of three main primitives: prompts, resources, and tool calling. The latter is particularly useful for what the speaker intends to discuss.
7. Tool calling is more complex than it might initially seem due to factors like dynamic tools, logging, sampling, tracing, and observability.
8. MCP allows tools to operate as subprocesses using standard in and standard out, which solves certain problems that an open API would not address.
9. The speaker shares a prototypical image of MCP, where an agent can connect to various tools without the need for pre-existing knowledge or design considerations.
10. Tools within the system can also act as agents, connecting to other tools over MCP or directly. This creates a challenge in managing LLM (Language Learning Model) access for each agent.
11. Sampling is a powerful feature of MCP that enables subagents to use an LLM from the original agent, reducing resource usage and management overhead.
12. Sampling is not yet widely supported but allows the server to make a request back through the client to the LLM for specific tasks.
13. The speaker will demonstrate sampling using Pantic AI, which has support for this feature.
14. An example use case presented is building a research agent to gather information about open-source packages or libraries.
15. One implemented tool in this example is making a BigQuery query to get download numbers for a specific package.
16. The speaker highlights the importance of providing type safety when accessing the MCP context within an agent validator or tool call.
17. Logging progress updates while the tool call is still executing provides value both for cursor-style agents and web application users.
18. The output from the query is formatted as XML, which is well-suited for LLM interpretation.
19. The speaker sets up an MCP server using FastAPI to register tools and pass user questions.
20. By performing inference inside a tool, context window overhead can be minimized, keeping the main agent's context lean.
21. The main application includes an agent definition that registers the Pippi MCP server as a tool for calling.
22. An example question is posed to the agent, asking how many downloads Pantic had this year, and the result is presented.
23. Observability of MCP calls, including the SQL query generated, can be viewed in Logfire or similar tools.
24. The speaker invites the audience to visit the Pantic booth for further discussion and demonstration.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!