ModelRefs / Tool Selection — AI Glossary

Tool Selection — AI Glossary

The agent's process of choosing the appropriate tool from its available toolkit to advance a task toward completion.

Overview

Tool selection is a core agent capability: given a task and a list of available tools (with descriptions and schemas), the LLM chooses which tool(s) to invoke and with what arguments. Quality depends on tool description clarity, the model's instruction-following ability, and the number of tools (large toolsets degrade selection accuracy).

Reference details

Topic	agents
Last reviewed	2026-06-24

How it works

Every tool's name, description and parameter schema is serialised into the prompt, so the model chooses while conditioned on all of them simultaneously.
That means context cost grows linearly with the catalogue and is paid on every request, whether or not any tool is eventually called.
Selection accuracy degrades as the catalogue grows because the descriptions compete for attention and near-duplicates become harder to separate, not because the model reasons worse.
A description is therefore a specification: it has to state what the tool covers and, where two tools are adjacent, where the boundary between them falls.
Retrieval over the tool descriptions is the structural fix — embed the request, select a handful of plausible tools, and offer only those — which converts a prompt that scales with the catalogue into a fixed-size one.
The selection is recorded in the call the model emits, so logging which tool was chosen alongside which tools were offered is what makes a wrong choice distinguishable from a tool that was never available.

Example: The boundary between two tools has to be written down

Give an agent `search_docs` (“search the documentation”) and `search_kb` (“search the knowledge base”) and it will pick between them roughly at random, because nothing in either description says which contains what. The model is not failing to reason; it has been handed an ambiguous specification. The fix is in the text: state the boundary explicitly — what each source contains, and when to prefer one — and selection accuracy moves without touching the model. Selection errors are usually specification bugs, and the first thing to try is rewriting the descriptions, not switching to a larger model.

Commonly confused with

Selection is choosing which tool; argument construction is filling it in correctly; execution is your code running it. They fail differently and need different fixes — a well-chosen tool with malformed arguments is a schema problem, not a selection one. Traces should record which tool was chosen and why it was available, or the two are indistinguishable after the fact.

When to use it

Reach for it when:

Any agent with more than two tools, where the choice is a real decision
As the first thing to instrument: log selections and review the wrong ones
Filtering the toolset per request when the catalogue is large, so only plausible tools are offered

Reach for something else when:

Attaching a large catalogue to every request — context cost and confusion both rise
Tools whose descriptions overlap, which makes correct selection impossible by construction
Assuming a bigger model fixes it before the descriptions have been rewritten

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Tool Selection — AI Glossary.

Frequently asked questions

What is Tool Selection?