How can you test your code when you don’t know what’s in it?

Summary

Testing Model Context Protocol (MCP) servers is uniquely challenging due to the non-deterministic nature of Large Language Models (LLMs) and agentic workflows. Because LLMs determine tool invocation sequences dynamically, traditional rigid testing methods are often ineffective, requiring a shift toward probabilistic verification and intent-based evaluation.

Key Points

• MCP servers define tools for AI agents, but the execution path is decided on the fly by the LLM rather than through a fixed, prescriptive workflow.
• Testing "named workflows" involves verifying a "skeleton" of tool invocations, ensuring a specific sequence of tools (e.g., Tool A followed by Tool B) occurs when provided with specific inputs.
• "Evals" (evaluations) represent an open-ended testing approach where an LLM is used to evaluate the output of another LLM.
• Developers should avoid "overfitting" to specific prompts or syntactical patterns, as updates to underlying models can render highly specific prompt engineering obsolete.
• Effective testing for agentic systems requires moving from low-level API and input/output shape validation to high-level validation of intent, functionality, and requirements.

Technical Details

Testing strategies for MCP servers generally fall into two categories: sequence verification and LLM-based evaluations. Sequence verification focuses on known or familiar workflows, where the test validates that a specific set of tool calls occurs in a certain order. This is particularly useful for critical paths, such as a routing step that decides between an open-ended path and a structured path like "create a new database record."

The second approach, evaluations (evals), is more open-ended and relies on using an LLM to judge the quality of an agent's output. This method is inherently probabilistic. Rather than seeking a "magical incantation" or a perfect prompt that produces a specific string, the goal is to design tests that ensure the model is likely to produce the correct output. This approach allows developers to "meet the model" rather than "beat the model," ensuring that as underlying LLMs improve, the agentic workflow benefits from increased intelligence without being constrained by overly restrictive, legacy prompt structures.

Impact / Why It Matters

Developers must transition from deterministic, syntax-based testing to probabilistic, intent-based frameworks to ensure the reliability of agentic workflows. Failure to move away from rigid prompt engineering can lead to brittle systems that break when underlying models are upgraded.