Tool Calls: Giving AI Agents the Keys

In a previous article we looked at how AI agents can plan tasks and do more than just answer questions. To perform those tasks, agents need a way to interact with other systems. This is where tool calls come in. Think of tool calls as the verbs that let an agent take action—fetching data, running a calculation, or sending a message.

What are tool calls?

A tool call is simply a request an agent makes to another service to get something done. This fits into the broader agent model discussed in our companion piece “From Chat to Action: Agentic AI in FileMaker”, where an agent isn’t limited to conversation but can plan, act and iterate. Imagine you ask a colleague to look up the weather: you give them a clear instruction and they return with the result. A tool call is the same idea, but formalised for an AI.

Developers are used to writing API calls, such as `GET /weather?city=Auckland`, to make software talk to a weather service. Agents don’t know how to write code. Instead, they select from a list of tool descriptions and fill in the blanks. Each description tells the agent what the tool does, what inputs it expects and how to provide them.

These descriptions are part of a growing set of AI‑native services. Unlike traditional APIs built for programmers, AI‑native services are written in plain language so that an agent can understand them. Once a tool is described in this standard way, any compliant agent can use it, no matter who wrote the tool.

How does a tool call work?

As Claris FileMaker developers, we can think of tool calls like calling a custom function or script step. You define the inputs and outputs, and the AI just plugs them together to carry out your request.

Tool calls follow a straightforward cycle:

1. Recognising the need for help – the agent reads your request and realises it doesn’t have the information or ability to answer on its own.
2. Picking a tool – it looks through its available tools and chooses the one that fits the job, guided by the descriptions.
3. Creating the request – the agent writes down the tool name and the details it needs, such as which city to look up.
4. Running the tool – the system running the agent executes the request. Importantly, the agent doesn’t run code itself; it hands off the work.
5. Getting the results – the tool sends back the answer.
6. Using the answer – the agent adds the new information to its conversation and plans what to do next.

Because the agent never runs code directly, you can set strict rules about which tools it may use and how their outputs are handled. For instance, you might only allow an agent to read data but require a human to approve any changes.

Why give AI access to tools?

Allowing an agent to call tools expands its usefulness well beyond writing sentences. In our companion article, From Chat to Action: Agentic AI in FileMaker, we discussed how a smart assistant can interpret your natural‑language query, plan what it needs to do, pick the right tool and iterate until it completes the task. When you marry that agentic logic with FileMaker’s new AI script steps, you get a system that understands what you mean and then works with your data in safe, predictable ways.

Here’s why tool access matters – and how it plays out in FileMaker:

1. Understanding your intent and run the right script step. The agent translates everyday language into specific actions. Using the Generate Response from Model script step (with appropriate tool definitions), a user can ask, “What is the average invoice total for customer X over their last 10 invoices?” The agent is able to interpret this request and call the appropriate tools to complete the task and answer your question.
2. Fetching fresh data from your own app. With the Generate Response from Model step in agentic mode, the AI can call FileMaker custom functions to run SQL queries. This means the agent isn’t guessing at answers; it’s retrieving up‑to‑date information directly from your database.
3. Chaining actions into a workflow. As described in the agent article, the AI can plan its next steps and decide what to do next. In terms of our question above, this means it can execute SQL to retrieve the last 10 invoice totals. It can then call the appropriate calculation tools to work out an average.
4. Keep things grounded and controlled. Tool calls rely on real data, which reduces the risk of hallucination. Because you define the tools – for example, which fields can be searched or which functions can be called – you control exactly what the agent can do. And with Set AI Call Logging, you can record every request and response for auditing.
5. Work seamlessly with structured data. Tools return results in predictable formats, making it easy for the agent to use them in its reasoning.

By combining natural‑language understanding with clearly defined FileMaker actions, agents become practical assistants that can interpret your queries and act on your behalf. The model isn’t free to do anything it likes – it calls only the tools you’ve described, and FileMaker’s script steps give you a high degree of control over those interactions.

Cautions when using tools

Giving an agent access to tools is powerful, but it also increases the risks. The same safeguards recommended for agents apply here:

• Scope – define exactly which tools are available and what actions they can perform.
• Transparency – keep logs of each call and its parameters. FileMaker’s tool call logging and other platforms’ tracing features are invaluable.
• Human oversight – review the outputs before applying changes. Use agents to suggest actions rather than performing them automatically.
• Security – keep API keys and credentials safe. Don’t expose them to the agent’s chat history.

You should also consider what happens if a tool returns no results or an error. Just like with FileMaker scripts, you should build in error handling so the agent doesn’t loop endlessly or give misleading answers.

Remember that tool calls are part of a bigger system. When tools are well described and access is carefully managed, they enable agents to move beyond static chat into practical assistants. Misconfigured or overly permissive tools, however, can lead to unwanted changes or security issues.

Summing up

Tool calls give AI the ability to act. By defining tools with clear descriptions and parameters, you create a safe environment where an agent can fetch fresh data, perform computations and automate tasks on your behalf. As the ecosystem of AI‑native services grows, understanding how to design and manage tool calls will help you unlock the full potential of agentic AI.

With these new commands in FileMaker we can get started now — so start small: define one or two safe tools, test them with the Generate Response from Model step, and log every call. You’ll quickly see how tool calls move AI from text "chat" towards practical, controlled automation.

Possible use case: Post-meeting voice capture in FileMaker Go

After a client meeting, a user can open FileMaker Go on their iPhone and say something like, “Create a new note for Customer X. The contents should be…”
The agent understands the request and then uses the appropriate tools to find the contact, create a new note record, and populate it with the dictated notes.
You could even get the AI to improve/summarise the dictated notes.

Tool calls turn AI from a talker into a doer. For Claris FileMaker developers, this means moving beyond chatbots into assistants that can fetch, calculate, and act inside your apps—with you always in control. Start experimenting with small, safe tools, and you’ll discover just how far agentic AI can take your FileMaker solutions.