Evaluation logs show that a GitHub Copilot agent is frequently calling a web search tool to look up information that already exists in the repository's internal documentation. This is causing slow, expensive runs. What tuning action should be taken?
Select an answer to reveal the explanation.
Short Explanation
If your agent keeps using web search to find information that already lives in your internal docs, you don't give it a bigger budget. You remove the wrong tool and give it a better one. Disabling web search and adding an MCP tool for internal documentation search fixes the behavior at the source.
Full Explanation
Tool access refinement is a key tuning lever when evaluation reveals that an agent is using the wrong tool for a given task. The issue here is that the agent has a general-purpose web search tool but needs a specialized internal documentation search capability.
The correct action is to disable web search and add an MCP tool that provides the agent access to internal documentation search. Disabling web search eliminates the costly misuse. Adding an MCP tool that indexes and searches internal documentation gives the agent a targeted, fast, and cost-effective way to find the information it needs. This is a direct tool access refinement: remove what isn't appropriate, add what is.
Increasing the agent's token budget to allow more web searches per run is wrong because it would make the problem more expensive, not less. The agent would make even more web searches. Budgets should be set after behavior is correct, not used to accommodate bad behavior.
Adding more detailed instructions telling the agent to avoid web search is wrong because instruction-based guidance is a soft control that may not be reliable. The agent may still choose web search when it thinks it's appropriate. A hard control, removing the tool, is more reliable than an instructional preference.
Switching from a tool-use architecture to a pure prompt-based architecture to eliminate tool misuse is wrong because it removes all tool capabilities, which would prevent the agent from doing its job at all. The problem is tool selection, not tool-use architecture.