Cut an agent's tool list down to the ones it actually uses

You are auditing the tool list of an AI agent that has become slow, expensive and unpredictable. Below are the tools it can call and a sample of recent runs.

Produce:

1. USAGE TABLE — Tool | Times called in the sample | What it was used for | Verdict (keep / merge / drop / gate behind a condition). One row per tool.
2. THE CASE FOR EACH DROP — one sentence per dropped tool naming the task that would break, and why that is acceptable.
3. OVERLAPS — pairs of tools that do nearly the same job, and which one to keep.
4. REPLACEMENT DESCRIPTIONS — for the tools you keep, rewrite each tool description in one line so the model picks correctly between them.
5. WHAT TO WATCH — 3 signals that dropping a tool was the wrong call.

Rules: no preamble, no praise for the current setup. Do not recommend adding tools. If the run sample is too small to judge a tool, say "insufficient evidence" rather than guessing.

TOOLS AND DESCRIPTIONS:
{{paste each tool name plus its current description}}

RECENT RUNS:
{{paste 5-20 runs showing which tools were called and the outcome}}