A developer adds explicit exclusions to two overlapping tool descriptions: search_users now says 'Do NOT use for subscription data — use get_subscription for that' and get_subscription now says 'Do NOT use for user profile data — use search_users for that.' Tests show the misbehavior rate dropped from 18% to under 1%. What technique did the developer apply?
Answer: Two-pass disambiguation — adding cross-references to both descriptions to make the boundary explicit from both sides
Two-pass disambiguation adds a cross-reference to both tools, creating a decision tree the model can follow without inference. When a query could match either tool, the model reads both descriptions, finds the exclusion in one, and routes to the other. This is more reliable than negative prompting alone (C) because it gives the model a positive alternative ('use get_subscription instead') rather than just saying what not to do. The tools were not deduplicated (A) or moved (D) — both remain in the agent's tool set. Cross-referencing is the specific disambiguation technique being applied.