Multiple Tools and Selection · Build AI Agents

Until now, the loop often had one tool. Real agents usually offer many.

You still own the loop. The new idea is simple: the model picks which tool to ask for. Your code runs the one it names.

On each call you send:

The person’s question
A list of tools the model may use

Example list:

Tool	What it does
`read_file`	Reads a file and returns the text
`get_time`	Returns the current time

The model does not see your code. It only sees each tool’s name, description, and inputs. Then it chooses.

You show a menu. The model picks.

  prompt + tool list
        │
        ▼
   Server / model
        │
        │  "use read_file" / "use get_time" / both / neither
        ▼
   Your program runs the named tool(s)

Who chooses?

The model chooses. Your loop does not say “always call read_file.”

It waits for the reply, then:

If there is no tool call → show the final answer
If there is a tool call → run that tool by name, send the result back, call again

Person asks…	Model might pick…
“What’s in notes.txt?”	`read_file`
“What time is it?”	`get_time`
“What’s in notes.txt, and what time is it?”	both (one after another, or both in one reply)

Clear names help the pick

Bad descriptions confuse the model.

Hard to choose	Easy to choose
`do_stuff` — “helps with files”	`read_file` — “Read a text file by path and return its contents.”
`info` — “returns info”	`get_time` — “Return the current local time as a short string.”

Write names and descriptions the way you would explain the tool to a new teammate.

Your code matches the name

When the model asks for a tool, you look at the name and run the matching function:

"read_file"  →  read the file
"get_time"   →  get the time
anything else →  send back an error (do not crash)

Only list tools you really built. If you advertise send_email but never wrote it, the loop will break when the model picks it.

Two tools in one reply

Sometimes one reply asks for two tools at once.

Then you:

Run both
Send both results back
Call the model again

Do not ignore one tool call. The model expects every result.

Example

Person: “What is in notes.txt, and what time is it?”

First reply might look like:

[
  { "tool": "read_file", "inputs": { "path": "notes.txt" } },
  { "tool": "get_time", "inputs": {} }
]

You run both. You add both results to history. Next call: the model writes the final answer.

Same loop as before — only now the menu has more than one item.

When things go wrong

Problem	Fix
Wrong tool picked	Make descriptions clearer; avoid two tools that sound the same
Needed tool never picked	Check it is on the list; make the question clearer
Unknown tool name	Return an error result; keep the loop going
Too many tools	Offer fewer tools for this job

Words to keep

Word	Meaning
Tool list	The tools you offer on this call
Selection	The model choosing which tool to ask for
Dispatcher	Your code that runs the right function for that name

See it in Code

The Code panel offers read_file and get_time, runs the loop, and prints which tool the model picked. Same idea in Python, Java, Go, and TypeScript.

A menu of tools