Tool Use/Function Calling - Tool Invocation Paradigm

Basic Information

• Type: Core Paradigm/Technical Standard for AI Agents
• First Appearance: 2023 (OpenAI Function Calling)
• Standardization Protocol: MCP (Model Context Protocol, Anthropic November 2024)
• Paper References: Toolformer (2023), ToolACE, etc.
• Current Status: Has become the core I/O layer for agent AI

Paradigm Description

Tool Use/Function Calling is the core paradigm that enables LLMs to interact with the external world. It allows models to output structured data (typically JSON) to instruct external systems to perform actions, rather than merely generating text. Traditional LLMs optimize for "the most persuasive next sentence," while agent-based LLMs optimize for "the most effective next action." Tool Use provides the I/O layer that connects these two worlds and is the foundation of all modern AI agent frameworks.

Core Mechanisms

• Natural Language → Structured Invocation: The model converts user intent into the JSON Schema required for API calls.
• Tool Definition: Developers define available tools and their parameters in JSON Schema format.
• Parameter Extraction: The model extracts the necessary parameters for tool invocation from the conversation context.
• Result Integration: The tool execution result is returned to the model, which integrates the result to generate a response.
• Multi-step Invocation: Supports chained invocations of multiple tools to complete complex tasks.
• Parallel Invocation: Some models support invoking multiple tools simultaneously.

Standardization Progress

• OpenAI Function Calling: Introduced in 2023, it has become the de facto industry standard.
• Anthropic Tool Use: Claude's implementation of tool invocation.
• Model Context Protocol (MCP): A standardization protocol introduced by Anthropic in November 2024, formally separating invocation logic from underlying implementation.
• MCPToolBench++: An MCP context tool evaluation benchmark released in August 2025.

Key Challenges

• Authentication and Security: Tool invocation requires secure authentication and permission management.
• API Compatibility: Integration between different API standards remains fragile.
• Hallucination Issues: Models may generate non-existent APIs or incorrect parameters.
• Complex Planning: The accuracy of planning for multi-step tool invocations is still below 50%.

Market Impact

• Tool invocation has become a key capability distinguishing "traditional LLMs" from "agent-based LLMs."
• The MCP protocol is becoming the unified standard for tool integration.
• Platforms like Composio have risen as a result.
• All mainstream LLMs (GPT-4, Claude, Gemini, etc.) now support tool invocation.

Relationship with the OpenClaw Ecosystem

Tool invocation is the core technology enabling OpenClaw personal AI agents to interact with the external world. OpenClaw should fully support the MCP protocol, allowing agents to invoke various tools and APIs through standardized interfaces. Additionally, OpenClaw can provide a tool definition marketplace, enabling users to share and reuse tool definitions, thereby reducing the cost of connecting agents to new services.