From Chat to Action: Agentic AI in FileMaker

Artificial intelligence is rapidly appearing everywhere, and new features arrive so quickly that it can be hard to separate marketing hype from genuine advances. When FileMaker 2024 added AI script steps, we explained how to harness those tools without getting lost in the hype. This article continues that mission and focuses on AI agents.

If you’re wondering what an agent does and how it differs from a typical chat interface, read on…

Simple Chat vs AI Agents: What’s The Difference?

A simple chat system is very good at having a conversation. You ask a question and it responds based on what it has learned from existing text. But that is all it can do. Imagine talking to a helpful librarian: they can tell you what they remember, but they aren’t allowed to leave their desk or look at any new books. They can’t go into the archive to check records, phone another library for an update or fill out a form for you.

Going Agentic

Making AI “agentic” can be likened to giving that librarian the keys to the archive, a telephone (or an internet connection!) and a checklist…

It can still talk to you, but can now also:

• Plan multiple steps – it breaks down a task into manageable steps rather than answering in one go.
• Use external tools – it knows when and how to run a search, query a database or use a calculator.
• Decide what to do next – based on what it finds, it chooses whether to keep searching, try a different approach or finish.
• Work iteratively – it can loop through these steps until it completes the task.

Put simply, a chat system is like a conversation that never leaves the room. An agent turns that conversation into a set of actions. It still "thinks” with the same kind of language model, but by calling tools and services it can fetch fresh information and carry out tasks that the model alone could never perform.

Why Not Stay With Simple Chat?

Traditional chat interfaces are wonderful for brainstorming and summarising text, but they have some important limits. Imagine reading from a reference book that was published last year: it will never include today’s news or weather. In the same way, language models are trained on data up to a certain date and do not know about anything that has happened since. They also can’t do anything with your data — they can’t look up a record in your database, check a live stock price or move a file.

Because a simple chat system cannot look up fresh information, it sometimes tries to fill the gap with a best guess. This is known as “hallucination”, when the system produces an answer that sounds plausible but isn’t correct. An agent overcomes these limits by recognising when it needs help, calling the appropriate tool and using the result to inform its answer. By tapping into live data sources and other services, the agent reduces the risk of outdated or invented answers and can actually perform tasks on your behalf.

AI Agents In Practice

When FileMaker added the Generate Response from Model script step, it included an Agentic Mode that behaves like an AI agent. In this mode, your script can send instructions that describe what the agent should do and which tools are available. The agent plans its workflow, calls the appropriate tool (e.g. `execute_sql` or a custom function) and returns the result. You can even inspect the order of tool calls for debugging purposes.

Similarly, the wider AI ecosystem is moving towards AI‑native services. Anthropic’s Model Context Protocol (MCP), introduced in late 2024, created a standard language for agents to talk to all sorts of services.

Software providers are building AI‑native interfaces that expose human‑readable descriptions and parameter schemas, making it easier for agents to understand how to interact. Within months of the protocol’s launch, thousands of tools appeared, allowing agents to access databases, browse the web, manipulate files and send emails.

For example, Xero introduced an MCP back in March this year, and other services such as Stripe, Paypal and Slack have MCP options.

What Should You Watch Out For?

Giving an agent access to tools is powerful, but it also increases the risks. Our earlier advice about not letting AI modify your data still applies. Keep these cautions in mind:

• Scope – define exactly which tools the agent can use. For example, expose a read‑only SQL query function but not a function that updates invoices.
• Transparency – log each tool call and its parameters. FileMaker’s tool call logging and other platforms’ tracing features are invaluable.
• Human oversight – review tool outputs before applying changes. Use agents to suggest actions rather than performing them automatically.
• Security – ensure that API keys or database credentials are stored securely and not exposed to the agent’s chat history.

Agents are also subject to the usual AI limitations: they inherit biases from their training data, can misinterpret instructions and may need careful prompting. While tool calls reduce hallucinations, they don’t eliminate them entirely.

Summing Up

An AI agent combines the conversational strengths of a language model with the ability to act. By defining what it can do and placing guardrails around those actions, you create a safe and structured environment where the agent can plan, call tools and iterate. This turns a chat‑style exchange into a workflow capable of retrieving real‑time information, performing computations reliably, and automating routine tasks.

With thoughtful design and human oversight, AI agents can move beyond simple chat into genuinely useful assistants that save time and reduce manual work. 

Example Use

Let's say you've come out of a meeting and want to capture your thoughts and actions

Using an AI agent you'd be able to open FileMaker Go 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 schedule the next action reminder or extract key points from what you've said.