# TRACER speeds up tool selection for LLM agents

> The new TRACER front‑end lets large language models pick tools with fewer tokens and higher accuracy, cutting token use by 80% in tests.

Oossa · 2026-10-16 · https://oossa.com/en/tracer-speeds-up-tool-selection-for-llm-agents

Researchers have released TRACER, a retrieval system that helps large language model agents choose the right tool from a large registry. The system stores past tool‑call sequences in a trie (a tree‑like index) and ranks possible next tools using a mix of data‑driven and context clues. In experiments, TRACER let GPT‑5.4 pick the correct tool 4.7 percentage points more often while showing only 5 of 128 tools, slashing input tokens per step by about 80%.

## How much better does it perform?

Across six test environments and four language models, TRACER beat the full‑registry approach in 21 out of 24 model‑environment pairs. Average gains in tool‑selection accuracy ranged from 6 to 10 points; the three cases where it lagged were no worse than 2.2 points. In a full‑turn autonomous run, the system matched GPT‑5.4’s overall task success while using 61% fewer input tokens per step.

## What this means for users

For developers building AI assistants that need to call external services—like checking inventory or activating a telecom plan—TRACER can make the assistant faster and cheaper to run, because it sends far fewer tokens to the language model. The lower token count also reduces the chance of hitting model context limits when many tools are available.

## The facts

- TRACER improves GPT‑5.4’s tool‑selection accuracy by 4.7 points in a mixed retail‑and‑telecom environment.
- It reduces input tokens per step by about 80%, exposing only 5 of 128 tools.
- In autonomous full‑turn execution, it matches task success while using 61% fewer tokens per step.
- Across six environments and four models, TRACER outperforms full‑registry selection in 21 of 24 model‑environment pairs.
- Average accuracy gains are between 6 and 10 points; the three losing cases fall short by at most 2.2 points.

## Why it matters

Developers can build smarter assistants that react faster and cost less to operate because they need to send fewer tokens to the language model. For end users, this could mean quicker responses when an AI needs to look up a price, book a service, or retrieve data from an external system.

## Sources & references

1. [TRACER: Trajectory-Aware Tool Retrieval for LLM Agents](https://arxiv.org/abs/2610.21267) – arXiv cs.IR, 2026-10-16

Last updated: 2026-10-16