Tool Use
Tool use is an agent's ability to call external tools when it needs information or capabilities beyond the model itself.
Prerequisites
What Is Tool Use?
A model's knowledge is frozen at training time and it has no way to affect the outside world on its own. Tool use is what closes that gap — an agent, given a set of available tools, decides when it needs one, which one to use, and what arguments to call it with, based on the goal and what's happened so far.
Agent Considers Options
Beyond the ModelThe model alone can't affect the outside world.
Select a Tool
Decides WhichBased on the goal and what has happened so far.
Call With Arguments
Specific InputsWhat exactly to pass to the chosen tool.
Read Result
New InformationWhat the tool actually returned.
Continue or Respond
Next DecisionUse another tool, or answer the user.
Key Idea
Tool use builds directly on function/tool calling — tool calling is the mechanism, tool use is the agent deciding when and which tool to invoke as part of pursuing a goal.
How an Agent Selects a Tool
Given a description of each available tool — usually its name, purpose, and expected inputs — the model decides which tool (if any) best matches what it needs right now. This decision quality depends heavily on how clearly each tool is described: a vague or overlapping tool description makes it harder for the model to choose correctly.
- Clear, distinct tool descriptions — reduce ambiguity about which tool applies to a given situation.
- Reasonable tool count — giving an agent access to dozens of overlapping tools tends to reduce selection accuracy, not improve capability.
- Validating tool results — a tool call can fail or return unexpected data; the agent (and the surrounding system) should handle that gracefully rather than assuming success.
A Real-World Example
An agent with access to both a "search internal docs" tool and a "search the web" tool has to decide, for a given question, which is more appropriate. A question about internal company policy should use the internal docs tool; a question about current public information should use the web search tool. Poor tool descriptions — or too many similar tools — make this choice less reliable, sometimes leading the agent to pick the wrong source entirely.
Common Mistakes
Giving an agent too many overlapping tools
A large or ambiguous toolset makes correct tool selection harder, not easier — fewer, clearly distinct tools usually perform better.
Writing vague tool descriptions
The model relies entirely on the description to decide when a tool applies — an unclear description leads to unreliable selection.
Assuming a tool call always succeeds
Tool calls can fail, time out, or return unexpected data — the agent needs a path to recognize and handle that.
Granting broad, unnecessary tool access
An agent should only have access to the tools it actually needs for its task — unnecessary access increases risk without adding value.
Not retrying or validating after a tool failure
Blindly continuing after a failed tool call can lead the agent to act on missing or incorrect information.
Interview Question
What is tool use, and how does an agent decide which tool to select?
Tool use is an agent's ability to call external tools when it needs information or capabilities the model doesn't have on its own — it builds directly on function or tool calling as the underlying mechanism. The agent selects a tool by matching the task at hand against each available tool's description, so tool selection quality depends heavily on how clear and distinct those descriptions are. Giving an agent too many overlapping tools tends to hurt selection accuracy rather than help, and the agent also needs to handle tool failures gracefully rather than assuming every call succeeds.
What an interviewer may ask next
- Why might giving an agent more tools actually hurt its performance?
- How does tool use relate to function calling?
- What should happen if a tool call fails or returns unexpected data?
Explain It in 30 Seconds
Tool use is an agent's ability to call external tools when it needs information or action beyond what the model can do alone, building on function or tool calling as the underlying mechanism. The agent selects a tool by matching its task against each tool's description, so clear, distinct descriptions matter a lot — too many overlapping tools tends to hurt selection accuracy. The agent also needs to handle a failed or unexpected tool result gracefully, not assume every call succeeds.