Tool-augmented LLM agents face a practical bottleneck: they need to find the right APIs from large registries, but cannot fit every tool definition into a single prompt. A new arXiv paper frames this as a question of ordering—should an agent plan first and then retrieve the tools the plan requires, or retrieve candidate tools first and then plan around them?
The study, described as a scale study of tool discovery architectures, notes that the dominant approach retrieves tools before planning execution. The available abstract is truncated, so the paper's specific findings and comparisons are not fully visible. What is clear is the motivation: finite context windows make exhaustive tool injection impossible, so agents must decide which tools are relevant, and when.
By posing the "plan or retrieve first" question, the work targets a core design choice for tool-augmented agents. The outcome could influence how future systems balance reasoning and external resource discovery.