Tool Call Engine

Understanding tool execution in @agent-tars/core through the MCP framework.

Overview

@agent-tars/core uses @tarko/agent for tool call execution, which provides different tool call engines:

• NativeToolCallEngine: Uses model's native function calling
• PromptEngineeringToolCallEngine: Prompt-based tool calling
• StructuredOutputsToolCallEngine: Structured output parsing

Tool Call Engines

The actual implementation is in multimodal/tarko/agent/src/tool-call-engine/:

Native Tool Calling

Uses the model's built-in function calling capabilities:

import { AgentTARS } from '@agent-tars/core';

const agent = new AgentTARS({
  model: {
    provider: 'openai',
    name: 'gpt-4'
  }
  // Uses NativeToolCallEngine by default for compatible models
});

Prompt Engineering

Falls back to prompt-based tool calling for models without native support:

const agent = new AgentTARS({
  model: {
    provider: 'custom',
    name: 'custom-model'
  }
  // Automatically uses PromptEngineeringToolCallEngine
});

Browser Tool Management

The main configuration for tool execution is through browser control strategies:

import { AgentTARS } from '@agent-tars/core';

const agent = new AgentTARS({
  browser: {
    control: 'hybrid' // 'hybrid', 'dom', 'visual-grounding'
  }
});

Browser Control Modes

• hybrid: Combines DOM and visual strategies
• dom: DOM-based browser control only
• visual-grounding: Vision-based control only

These are implemented through BrowserToolsManager and strategy pattern.

Tool Execution Flow

Tools are executed through the MCP framework:

1. Tool Registration: Tools are registered via MCP servers
2. Tool Discovery: Agent discovers available tools
3. Tool Calling: Model decides which tools to call
4. Tool Execution: MCP framework executes the tool
5. Result Processing: Results are returned to the model

Next Steps

• Learn about Tool Development to create custom tools
• Explore Tool Management for organizing and filtering tools