Add the think tool

Create the built-in tool and register it with the other tools available to the agent.

Basic setup

import { AgentBuilder, createThinkTool } from '@anvia/core'

const agent = new AgentBuilder('support-investigator', model)
  .instructions(
    'Use think when you need to compare multiple tool results before answering.',
  )
  .tools([
    createThinkTool(),
    searchTicketsTool,
    getAccountTool,
  ])
  .defaultMaxTurns(6)
  .build()

The model decides whether to call the tool. Registering it does not make every request use it, so pair it with a specific instruction.

Customize the definition

The default name is think. Change it if that name already exists or a more precise capability name will guide the model better:

const evidenceCheckTool = createThinkTool({
  name: 'review_evidence',
  description:
    'Record a concise assessment of the evidence before choosing the next action.',
})

Option	Default	Purpose
name	think	Model-facing tool name.
description	Built-in description	Explains when the model should call it.

The input and output contract does not change when the tool is renamed:

type ThinkInput = {
  thought: string
}

type ThinkOutput = string

Add it to the right agents

Do not add think globally merely because an agent can call tools. Attach it to agents whose work actually involves investigation, comparison, or multi-step decisions.

For a large dynamically retrieved catalog, keep think static so it is always available during the run:

const agent = new AgentBuilder('operations', model)
  .tools([createThinkTool()])
  .dynamicTools(operationsToolIndex, {
    topK: 5,
    threshold: 0.72,
  })
  .build()

Static registration makes the checkpoint available on every turn while operational tools are selected from the index.