Preprint: Learning Tool-Output Projections from Agent Traces Cuts Billed Tokens up to 56%

solyarisoftware · x · 2026-11-10

Independent researcher Alessio Pomaro released the preprint Agents Read More Than They Use plus the open-source tool tracetrim: LLM agents read tool outputs written for programs, visibly reuse only a fraction of the fields, yet pay for all of it every turn.

The approach learns from an agent's own execution traces which fields of each tool output it visibly reuses, and trims the rest.

Tested on five agents (three real APIs, two simulated OpenAPI servers): a projection learned from seven runs removed 63%/70% of tool-output tokens on two real-API agents and 69%/77% on simulated ones; compact JSON without removing fields already saves 13%-17%, a full minifier up to 33%.
With an expand tool for hidden fields, billed input tokens fell 56% and 32% in live runs, all passing gates — though a re-check found one unsupported value among 180 keyword-volume pairs.
Failure modes: values transformed before writing, reasoning-only fields, and agents inventing identifiers when none is provided.
Learning only from tool calls (gateway view) missed a used field in 49% of trimmed structured calls, vs 4% when learning from complete traces.

Original post →

Preprint: Learning Tool-Output Projections from Agent Traces Cuts Billed Tokens up to 56%

More from coding & agent