LLM agents often pick tools from large registries, but showing the entire registry to the model is expensive in tokens and can make selection worse as the registry scales. That is the problem TRACER targets, according to its arXiv abstract.
TRACER is described as a retrieval front-end that indexes observed tool-call sequences. Instead of exposing every tool, it uses trajectory information to retrieve only the relevant ones, potentially cutting token costs and improving selection accuracy.
The abstract is brief and does not include implementation details or evaluation results, so the claims are limited to the proposed design. No other sources are available for comparison, so this summary reflects only the single arXiv posting.