Tool Use
An LLM's ability to call external functions or APIs as part of its response, enabling actions beyond text generation. Also called function calling.
Also known as: function calling · API calling · actions
What is Tool Use?
Tool use is an LLM's ability to call external functions or APIs during its response, allowing the model to read databases, send messages, or perform actions in the real world. Also called function calling. Tool use is the architectural feature that distinguishes AI agents from chatbots: a chatbot only generates text; an agent uses tools to do work. Production tool use requires three things: a tool catalog (what the agent can call), structured arguments (what each call needs), and bounded permissions (what the agent is allowed to do versus must escalate).
Christos Papadimitriou, theagency47 · Updated May 15, 2026Why tool use matters
Without tool use, an LLM is a text-in / text-out system. With it, the LLM can read databases, write records, send emails, schedule meetings, check inventory, post to Slack, anything that has an API.
In 2026, production-grade tool use is what makes AI agents viable for business work. The 2023 “chatbot wave” largely lacked tool use; the 2024-2026 “agent wave” is built around it.
How theagency47 uses tools in agent builds
A typical agent has 3–8 tools defined in its catalog. For example, a customer support agent (like Yiannis) might have:
lookup_order(order_id), reads the order databaseget_customer_profile(email), reads customer historyissue_refund(order_id, amount, reason), bounded refund authorityescalate_to_human(ticket_id, reason, suggested_response), escalation pathupdate_ticket_status(ticket_id, status), writes ticket state
Each tool has a JSON schema for arguments, a security boundary, and an audit log entry on every call.
FAQ
Are tools different per agent?
Yes. Each agent has a tool catalog tailored to its job. Sales agents call CRM tools; support agents call order DB tools; email agents call Gmail tools. Cross-tool access is restricted unless explicitly granted.
How are dangerous tools handled?
By bounded autonomy: high-stakes tools (delete data, send external email, large refunds) require human approval or are excluded entirely. The eval suite tests this discipline before launch.
Related terms
AI Agent
A software system that takes a goal, plans steps, calls tools, and acts autonomously to produce an output, without step-by-step human prompting.
Bounded Autonomy
The engineering pattern of giving an AI agent specific permissioned actions and explicit escalation rules, letting it act independently within tight, defined boundaries.
Escalation Rule
A configured condition that hands an agent task off to a human when the agent is uncertain, out of scope, or facing high-stakes input.