AI agent tool use: How it works and practical examples

AI agent tool use is the process of selecting and using an external capability to get information or perform an action that the AI model can’t handle on its own, then using the result to determine what to do next.

The agent doesn’t directly perform the external action. If it needs to search the web, send an email, or query a database, it generates a request describing what needs to be done.

The system running the agent forwards the request to the appropriate tool, which performs the action and returns the result.

Agents can choose between available tools, combine several of them in one task, and use mechanisms such as function calling and MCP to interact with external systems.

The process follows a simple loop: the agent chooses a tool, sends it a request, receives the result, and decides what to do next. AI agents can repeat this process several times to complete a task.

Suppose you ask an agent to find a meeting time when you and two colleagues are available. The process would work like this:

1. Identify the need for a tool. The agent determines that the task requires information or an action that the AI model can’t provide on its own. To find a meeting time, for example, it needs access to current calendar data.
2. Generate a tool call. The model produces a structured request for the selected tool rather than performing the action itself. The request identifies the tool and supplies the arguments it needs, such as the attendees, date range, working hours, and meeting duration.
3. Pass the request to the tool. The system running the agent receives the tool call and sends the supplied arguments to the calendar tool. This separates the model’s request from the external system that actually performs the action.
4. Execute the requested action. The calendar tool uses its access to the external system to check the attendees’ calendars for matching time slots. What it can access or change depends on the permissions granted to that tool.
5. Return the result to the agent. The tool sends its output back through the system running the agent. In this case, the result might contain several time slots when all three attendees are available.
6. Use the result to continue the task. The returned data becomes new information that the agent can work with. It can present the available times to the user or determine that another tool call is required, such as searching a wider date range or creating an event after the user chooses a time.

Some tools give AI agents access to external information, while others let them run code, work with files, interact with software, or perform actions in business applications.

The most common tools AI agents use include:

• Web search tools let agents find current information on the internet. An agent can use them to look up recent news, product information, documentation, or other web content needed to complete a task.
• Retrieval and database tools give agents access to stored information. They can retrieve customer records from a database, find relevant passages in a knowledge base, or pull information from internal documents when a task depends on data the model doesn’t already have.
• Code and computation tools let agents run code and perform calculations. They can analyze a dataset, transform data into another format, calculate values, or run a script and use its output to continue a task.
• File tools allow agents to work directly with files. Depending on the available operations, an agent can read a document, extract information from it, create a new file, edit existing content, rename files, or organize them into folders.
• Browser and computer-use tools allow agents to interact with software through its user interface. An agent can navigate a website, click buttons, fill in forms, or work with an application when the required task needs interaction with its interface.
• Business tools connect agents with applications used for day-to-day work, including email, calendars, CRMs, messaging platforms, and project management systems. These tools can let an agent check a calendar before scheduling a meeting, retrieve a customer record, send a message, or update a project task.

An AI agent chooses from the tools available to it based on the task it needs to complete. To choose the right one, it relies on information about each tool, including its name, description, and required inputs, as well as what it already knows about the current task.

AI agent reasoning typically follows this path:

1. Identifying what the task requires. The agent uses the current goal and task context to determine its next need. If it is preparing for a customer meeting, for example, it might first need to retrieve the customer’s account information.
2. Finding a tool that can meet that need. The agent compares the available tools using their names and descriptions. A tool called search_customer_records with a description saying it retrieves CRM records, is a better match for this step than a calendar or email tool.
3. Checking whether it can provide the required inputs. The tool’s input schema tells the agent what information to include and how to format it. If a CRM lookup requires a customer email address, the agent takes that email from the current task and includes it in the request.
4. Calling the tool and evaluating the result. After the tool runs, the agent checks what it returned and whether that result satisfies the current need. A CRM lookup might return the customer’s company name, account details, and recent activity.
5. Deciding whether another tool is needed. The result can create the next step. With the company name from the CRM, the agent could search internal documents for relevant project information, then use a calendar tool to retrieve the meeting time and attendees. If it already has everything required to complete the task, it can stop instead.

Tool use is the overall process of an AI agent selecting and using external capabilities to complete a task. Function calling is one way the model can request a specific action by generating the inputs a predefined function requires, while MCP provides a standard for making external tools and data sources available to AI applications.

AI agent skills serve a different purpose: they provide reusable instructions and resources that guide an agent in completing a task.

The table below compares what each concept means and the role it plays when an AI agent works with external tools:

For example, an agent preparing a weekly sales report could use an agent skill that includes instructions on what data to collect and how to structure the report.

Tool use allows it to retrieve the required data from a CRM, function calling can define the request sent to the CRM tool, and MCP can provide a standardized way for the AI application to access that tool.

AI agent tool use can support tasks such as researching information, following up with sales prospects, fixing code, and resolving customer support requests.

In each case, the agent can use different tools as the task progresses, with one tool’s result helping determine what information or action is needed next.

Research agent

A research agent can combine search, file retrieval, and data analysis tools when a question can’t be answered from a single source. Suppose it’s asked to compare how several cloud providers changed their prices over the past year.

The agent could use a web search tool to find current pricing pages and recent announcements, and then retrieve existing reports or spreadsheets containing earlier pricing data.

If the information needs to be compared across dozens of products, it could use a code tool to clean the data, calculate percentage changes, and organize the results.

Each tool handles a different part of the research process. Search provides current information, file tools provide existing source material, and code handles analysis that would be difficult to perform reliably from text alone.

The agent can then use the combined results to prepare its comparison and identify where information is missing or needs further verification.

Sales agent

A sales agent can use CRM and communication tools to complete a follow-up based on the latest information about a prospect.

For example, after a sales call, it could retrieve the prospect’s CRM record to check the deal stage, previous interactions, assigned salesperson, and notes from earlier conversations.

The CRM result provides the agent with context for deciding what the follow-up should include. It could then prepare an email based on the agreed next steps and use an email tool to send it to the prospect.

Once the message is sent, the agent could return to the CRM to record the interaction, update the deal stage, or create a follow-up task for the salesperson.

Here, tool use allows the agent to carry information from one system into the next action rather than treating the email, CRM update, and follow-up task as separate requests.

Business operations agent

A business operations agent can use tools from different parts of a business to complete tasks that would otherwise require switching between several apps.

For instance, it might use a calendar tool to check upcoming meetings, retrieve related information from connected files or business apps, and use an email tool to send an update. If the task needs to run again later, the agent could also schedule it as a recurring job.

Hostinger Agent supports this type of tool use by working with Hostinger services and third-party apps such as Gmail, Outlook, HubSpot, Notion, Google Calendar, Google Drive, and Jira.

It can also search the web, work with uploaded files, and run code in an isolated sandbox, allowing it to combine different tools as a task progresses.

Coding agent

A coding agent can use file and code-execution tools to make a change and verify whether it actually works.

If it’s asked to fix a bug, it could first inspect the relevant project files and search the codebase for the functions connected to the reported problem.

After identifying a likely cause, the agent could edit the affected file and use a code-execution tool to run the project’s tests.

A failed test becomes new information that the agent can act on. It can inspect the error message, trace it back to the relevant code, make another change, and run the tests again.

This creates a feedback loop between editing and testing. Rather than assuming that a code change solved the problem, the agent uses the output from its tools to check the result and determine whether further work is needed.

Customer support agent

A customer support agent can use tools to gather information from several systems before taking action on a ticket.

Say a customer reports that an order hasn’t arrived. The agent could retrieve the support ticket and account details, then use an order-management tool to find the purchase and a shipping tool to check its current delivery status.

Those results determine what happens next. If the package is still in transit, the agent could send the customer the latest tracking information.

If the carrier reports it as lost, the agent could follow the appropriate support procedure, update the ticket, and route the case for a replacement or refund when human approval is required.

As with other AI agent examples, the agent isn’t simply calling several tools in sequence. It uses the information returned by each tool to decide which tool or action is appropriate next, allowing the workflow to change according to the customer’s actual situation.

Reliable AI agent tool use depends on giving agents clearly defined tools, limiting what those tools can access and change, validating requests and results, and adding safeguards around higher-risk actions.

These practices reduce incorrect tool calls while limiting the impact of errors or malicious content returned by external systems.

To make tool use more reliable and safer, follow these practices:

• Keep toolsets small and clearly scoped. Give the agent only the tools relevant to the tasks it needs to perform. If several tools perform similar actions or the agent has access to capabilities it doesn’t need, choosing the appropriate tool becomes harder, and unnecessary systems or data may get exposed.
• Use precise tool names, descriptions, and schemas. Make it clear what each tool does, when the agent should use it, and which inputs it accepts. Instead of a vague tool such as manage_customer, separate capabilities such as get_customer_record and update_customer_record, with required inputs and accepted values defined in their schemas. This helps the agent select the appropriate action and construct a request that the tool can execute.
• Apply least-privilege access. Limit each tool to the permissions and data required for its intended task. An agent that only needs to check calendar availability, for instance, can be given permission to read events without permission to create, edit, or delete them. If the agent or tool behaves unexpectedly, the available access limits what it can affect.
• Validate tool calls and results. Check arguments before sending a tool call for execution, particularly when they control important actions. A payment tool could verify the amount, currency, and recipient before processing a transaction. Returned data should also be checked for the expected format and values before the agent relies on it in another step.
• Return useful errors. Tell the agent why a tool call failed, rather than returning a generic error. Identifying a missing argument, an expired credential, a rate limit, or an unavailable service gives the agent information to correct the request, try again later, choose another appropriate action, or stop.
• Log tool activity. Keep a record of which tool was used, when it ran, who or what initiated the action, the arguments supplied where appropriate, and the result or error it returned. These records make it easier to investigate unexpected actions, troubleshoot failed agentic workflows, and audit how an agent used its available tools.
• Require confirmation for irreversible or sensitive actions. Add human approval before allowing an agent to delete data, publish content, send sensitive communications, change permissions, make purchases, or perform other actions with significant consequences. The confirmation should clearly show what the agent intends to do, so the user can review the action before it is executed.
• Treat tool output as potentially untrusted. Web pages, retrieved documents, emails, database records, and other information returned by tools can contain misleading or malicious instructions. Treat this content as data needed for the task, not as instructions that can change the agent’s goals, expand its permissions, or authorize additional actions.

Once an agent can select and use tools, those capabilities need to be connected to the external systems where the information and actions are available.

Tool use describes how an agent chooses and uses a capability during a task, while AI agent integrations provide the connections to APIs, business applications, databases, and other external systems that make many of those capabilities available.

These integrations also control what the agent can access and do in each connected system. Authentication establishes the connection, while permissions can limit access to specific data and actions, such as reading records but not modifying or deleting them.

All of the tutorial content on this website is subject to

Hostinger’s rigorous editorial standards and values.