Reading - Agent Tools & Interoperability with MCP
Reading - Agent Tools & Interoperability with MCP1.
What's a tool
A tool is an external function or service that a language model can invoke at runtime to perform actions beyond its static training data.
OpenAI introduced function calling (circa mid-2023), which allows models to invoke external code. Modern tools are built on top of this function-calling capability.
Tool definition
A tool definition is a contract between the model and the client. This contract must include:
- a clear, unambiguous function name
- a descriptive docstring explaining what the tool does and when it should be used
- a strictly typed schema for input arguments
Example:
N × M Integration Problem
Consider a system with N AI models and M tools. Without standardization, each model requires a custom integration for each tool—resulting in N × M point-to-point connectors.
This quickly becomes unmanageable as both models and tools scale.
LSP
Before LSP, each code editor needed a separate plugin for each programming language. With 10 editors and 10 languages, that meant 100 plugins.
Microsoft introduced the Language Server Protocol (LSP). It provided:
- a client–server architecture with a JSON-RPC–based message interface
- a language server that implements language-specific logic (parsing, diagnostics, completions)
- servers advertising the capabilities they support
- clients connecting to servers and consuming those capabilities
MCP
MCP follows a similar architectural pattern to LSP. It provides:
- an MCP host, which advertises the tools it exposes
- communication via JSON-RPC 2.0
- a standardized way for clients to discover and invoke tools
Tool discovery example:
Tool Invocation
User query: “What’s the weather in New York?”
The LLM decides to call get_weather(city="NY").
The server executes the function and returns the result, which the client feeds back into the model’s context.
Sampling
Sometimes a tool needs LLM assistance to complete its task—for example, summarizing a large document before returning a result.
Flow:
- MCP server receives a request (e.g.,
fetch_and_summarize_pdf) - it retrieves the raw PDF
- instead of processing it locally, it issues a sampling request back to the client: “Please summarize this 50-page document using your LLM.”
- the client runs the summarization and returns the result to the server
This keeps the server lightweight while still leveraging model intelligence.
Elicitation (Tool → Human)
When a tool lacks critical information—such as confirmation for a destructive action—it can elicit input from a human operator.
Example:
- a tool attempts to delete a file
- deletion requires a 2FA code or explicit confirmation
- the tool pauses execution and requests human input
Best Practices for Designing Agent Tools
Describe intent, not implementation Prefer: “Create a high-priority bug ticket in Jira” Avoid: “POST to /rest/api/3/issue with JSON payload…”
Granularity matters Each tool should perform one atomic action. Compose complex workflows by chaining simple tools.
Minimize output payloads Never return megabytes of raw data to the LLM. If output is large:
- store it externally (object storage, cache)
- return a reference instead (URL, ID, hash)
Validate inputs strictly Enforce type and domain constraints (e.g., brightness ∈ [0,100]) to prevent runtime errors.
Idempotency & safety Design tools to be safe on retries (e.g., idempotency keys for financial or destructive operations).
References
comments powered by Disqus