Tool Calling

Function calling (tool calling) refers to the ability of a large language model to proactively request the invocation of external functions or APIs during the generation of a response.

The core distinction between agentic coding tools and traditional code completion tools lies in autonomy and task scope; agentic tools can accept high-level instructions and autonomously plan steps, invoke tools, and execute iteratively.

Evaluation dimensions for the tool-calling layer include whether the Agent selects the correct tool, whether it passes correctly formatted and semantically reasonable parameters, and whether it can correctly parse the results returned by the tool

The core challenge of agentic programming lies in the model's need to plan multi-step actions, handle error feedback during execution, and coordinate operations across different files and tools.

During inference, large models read tool metadata as context and use semantic understanding to determine whether a tool needs to be called and what parameters to pass, a process known as Function Calling or Tool Use.

Agent mode is implemented through a 'tool-calling' mechanism, whereby the model can autonomously decide to call external APIs, execute shell commands, read and write files, or browse the web.

In the tool-calling mechanism, developers attach a tool description list in JSON Schema format when calling the LLM API, and the model outputs call parameters in structured JSON format

工具调用（Function Calling/Tool Use）是Agent系统的核心能力，允许模型动态调用搜索引擎、代码解释器、数据库查询、API接口等外部资源

Large models are essentially just a brain with no ability to obtain external information, and they need to acquire real-time data through tool calls.

工具调用对话的序列化需要存储包含多种角色（user、assistant、tool）和多种消息类型（文本、工具调用请求、工具调用结果）的复杂序列

Tool calling refers to the mechanism by which an Agent requests the execution of an external function through a structured message and has the result returned; the returned content enters the conversation history as a message with role: 'tool' and from then on permanently occupies the context.

In TokenRouter, the judge must be a model that supports tool calling, because the verdict is returned via a tool call, and the model selector only lists such models.

Function calling and local sandboxes are the technical foundation for an Agent to operate in a local environment: the model outputs structured call requests, which are executed by the execution layer in the device or cloud sandbox, and the results are then passed back to the model.

Tool calling and structured output are two different model capabilities, and one cannot assume that a coding model can drive a particular harness.

When selecting a model in God, you should confirm that it is marked as supporting tool calling, and first verify it once with a small task like "find the most recent customer email and identify the subject and date".

Tool Calling is a key capability for whether a local model can handle email summarization; a model that supports tool calling can proactively call external functions (such as reading mailbox emails) according to instructions, while one that does not can only answer out of thin air or rely on already injected context, potentially producing summaries that do not match the real emails.

God supports a Tool Calling mechanism, enabling the model to proactively issue structured requests during its response, such as querying calendar APIs, reading notes, or retrieving history, and to integrate real data.

Ordinary conversational models can only generate text, whereas models supporting tool calling can proactively issue structured requests, such as querying calendar APIs, reading saved notes, or retrieving history.

Tool calling is a separate skill trained separately from writing code, and below a certain parameter threshold models struggle to emit valid tool calls reliably.

The model never directly touches databases, APIs, or files; it issues structured tool requests, which the orchestrator validates, executes, and returns results for, and the model can only reach the outside world through the capabilities the application chooses to expose.