How to Reduce Tokens in Agent Tool Responses

Set a response budget

Give every tool a default result count and maximum serialized size. A search tool might return five passages with IDs, titles and snippets; a document tool can fetch one full item only when needed.

Projection matters. Database objects contain audit fields, nested objects and repeated labels that are essential internally but irrelevant to the next model decision.

Use progressive disclosure

search_workspace(query, filters, limit)
  → ranked snippets and document IDs

get_document(document_id)
  → one selected document and provenance

This pattern spends context only on promising material and records which candidates were shown and which item was opened.

Compress without losing signal

• Prefer stable keys and arrays over repeated prose.
• Return one canonical timestamp.
• Truncate at meaningful boundaries and mark truncation.
• Keep source IDs and titles.
• Never compress away permission or error information.

SPYN3 REST responses default to a compact TOON representation and can be requested as JSON. The principle is to preserve decision-relevant structure while removing repetition.

Evaluate the whole task

A response that is 50% smaller but causes extra calls or missed evidence may cost more overall. Compare successful completion, total calls, total tokens, latency and citation correctness. Use retrieval evaluation to ensure aggressive limits do not remove required evidence.