Agent tool routing

Choose which tool an assistant should use next, from a fixed list.

Problem

Letting a generative model choose tools freely makes behaviour hard to predict and test. A wrong tool call can expose data or take an action nobody asked for.

Goal

Turn the next step into a typed choice that code can check, log and gate, and ask a clarifying question when the choice is unclear.

Input / state

The latest message, a short summary of the conversation, and the available tools.

latest_message: The user's most recent message.
conversation_summary: What has happened so far, in a few sentences.
available_tools: The tools the signed-in user is allowed to use.

Decision question

ChoiceQuestion key next_tool

Which tool should the support assistant use next to handle `latest_message`?

Options

search_help_centerThe user asks how something works or how to do something the help center documents.
lookup_orderThe user asks about the status, contents or delivery of a specific order.
create_ticketThe user reports a problem that needs a person to investigate or act.
no_toolA greeting or thanks, or something that can be answered from the conversation so far.

Allowed outcomes

Decided in your code from the answer and its confidence. Thresholds are yours to set and test.

Search help centerChoice is search_help_center and confidence meets your threshold.
Look up orderChoice is lookup_order and confidence meets your threshold.
Create ticketChoice is create_ticket and confidence meets your threshold.
Reply directlyChoice is no_tool and confidence meets your threshold.
Ask to clarifyConfidence is below your threshold.

Example scenarios

Predefined demonstrations. No model is called; the outcome shown is an illustrative judgment.

Example state

{
  "latest_message": "How do I export my data to CSV?",
  "conversation_summary": "New conversation.",
  "available_tools": [
    "search_help_center",
    "lookup_order",
    "create_ticket"
  ]
}

Outcome

Illustrative

Search help center

Choice is search_help_center and confidence meets your threshold.

Why

A documented how-to question.

Better together with

Complementary classifiers, and what each one adds here.

Prompt injection signalNoul
Screen the message before a tool is chosen, and restrict tools when the signal is unclear.
Combined inGrounded support assistant turn
Retrieved passage relevanceScore
When the router picks search, score what comes back before it reaches the answer.
Combined inGrounded support assistant turn
In-scope request checkNoul
Lets the assistant tell when no tool fits an off-topic request.

Used in chains

Grounded support assistant turnStep 2 of 4

Sources

Primary documentation this classifier builds on. Scenarios on this page are illustrative, not taken from production use.

SourcedTypeSafe docs · Cookbook
Function calling
Maps a request to a function name and closed-set arguments with Choice questions, each with a confidence.
SourcedTypeSafe docs · Pattern
Companion Noul
Pairs a Choice with a Noul asking whether any option fits, because a Choice always picks an option even when none applies.

Using it well

Arguments that take values from a fixed list can be their own Choice questions. Free-text arguments, such as an order ID, should be extracted and validated in code.
Offer only the tools the signed-in user may use, and enforce permissions in the tool, not in the question.
Require explicit confirmation for tools with side effects, such as refunds or cancellations.

Request definition

POST https://api.typesafe.ai/v1/systemone

{
  "model": "jev-latest",
  "state": {
    "latest_message": "<latest_message>",
    "conversation_summary": "<conversation_summary>",
    "available_tools": "<available_tools>"
  },
  "questions": {
    "next_tool": {
      "type": "choice",
      "instructions": "Which tool should the support assistant use next to handle `latest_message`?",
      "criteria": {
        "search_help_center": "The user asks how something works or how to do something the help center documents.",
        "lookup_order": "The user asks about the status, contents or delivery of a specific order.",
        "create_ticket": "The user reports a problem that needs a person to investigate or act.",
        "no_tool": "A greeting or thanks, or something that can be answered from the conversation so far."
      }
    }
  }
}

Request shape from the System One API reference; a Kev server accepts the same body. Replace the placeholder state with your data. Requests need your own API key or server.