How to expose tools to AI agents without clogging their context

Описание и применение

Внедрение поиска инструментов с закрепленными часто используемыми инструментами сократило потребление токенов на 69 %, улучшив эффективность AI‑агентов.

Задача

AI agents struggle to manage large tool catalogs because tool definitions, selection, and results consume excessive context tokens, hindering performance and clarity.

Результат

Implementing discovery with pinned tools cut input token usage by 69% compared to loading all tool definitions, improving efficiency and reducing context clutter.

Рабочий процесс

• Configure Agent Sessions to expose a small working set of tools and enable discovery for the rest.
• Pin frequently used tools (e.g., meeting and account actions) in the session configuration.
• When a task starts, load only the pinned tools into the model’s context.
• If the agent needs an additional tool, invoke the meta tool nango_tool_search to find matching definitions.
• Execute the selected tool via nango_execute or directly if already loaded.
• Enforce execution permissions at the backend, ensuring the agent cannot broaden its scope.
• Measure token usage for the entire task to decide which tools to pin.
• Iterate on pinning strategy based on observed usage patterns.

Доказательства

Discovery with two pinned tools used 69% fewer input tokens per task than loading all 57 tool definitions upfront in our synthetic test.

Источник

Tool exposure creates three separate costs: Definitions, Selection, Results.

Источник

Separate tool permissions from tool visibility. User authorization, session permissions, and tool visibility control different things.

Источник