Back to Glossary
Function Calling UX
Synonyms: tool use UX, tool call UI, generative UI, agent action UI
Do not index
Definition
Function calling is when an AI model asks your app to run a specific function, such as
search_flights or create_invoice, fills in the inputs, and gets the result back. Function calling UX decides what the user sees at each step: which tool runs, with what inputs, how long it takes, what came back, and what to do when it fails.Use cases
Invisible tool calls leave users guessing: working or frozen, done or broken.
- The silent spinner: An agent makes 4 tool calls behind one "Thinking..." label. After 40 seconds the user refreshes and kills the run.
- JSON in the chat: The result shows up as
{"status":"ok","id":"inv_8812"}. Did the invoice go out? No idea.
- Wrong inputs, no preview: The model books Portland, Oregon. The user meant Portland, Maine, and never saw the inputs before the call ran.
How it's used in practice
- Name each call in plain words: Give each function a label, e.g. "Searching flights to Lisbon", shown the moment the call starts.
- Preview inputs before risky calls: "Book 2 seats, 14 Nov, $412" with Edit and Confirm.
- Render results as components: A flight card or table, built per tool.
- Design 4 states per tool card: Running, done, needs input, failed.
- Report partial success: "3 of 4 suppliers replied. Acme's API timed out." Name the failing service.
Put the tool names and descriptions through UX writing review. The model picks tools by reading them. Vague names such as
search and search_v2 cause wrong calls that show up as UI bugs in testing.Challenges & limitations
- Chains get noisy: One request can fire 10+ calls. Listing all of them buries the answer, and collapsing them hides what happened.
- Latency isn't yours: Third-party APIs set the wait. The UI can only report it.
- Designers rarely see the schema: Tool names and error formats get decided in engineering tickets.
Free resources
- OpenAI: Function calling guide — how models call your functions with JSON Schema inputs, including strict mode and multiple calls per response.
- Google AI for Developers: Function calling with the Gemini API — function declarations plus the auto, any, none, and validated calling modes.
- Smashing Magazine: Practical Interface Patterns For AI Transparency (Part 2) — Victor Yocco's step checklist, partial-success, and "disentangling the tool" patterns.
