Tool Calling
Core Facts
Function Calling (Tool Use) is a standardized mechanism for large models to invoke external tools, where developers pre-declare tool names, descriptions, and parameter schemas, and the model autonomously determines when to make a call.
The underlying mechanism of Tool Use is that developers pre-define the name, description, and parameter structure (Schema) of tools, and the model determines during inference when and which tool to call and outputs structured invocation instructions.
The tool-calling (Function Calling) mechanism allows the model to invoke external tools during reasoning, such as file read/write operations, terminal commands, and search APIs.
Timeline (last 90 days)
There are three common tool invocation approaches: full injection, rule-based routing, and embedding-based retrieval.
Small-parameter models easily confuse similar tools when given long tool lists; in practice the number of tools exposed at once should be limited and parameters described specifically.
The tool-calling mechanism allows a large language model to output a structured call request (usually JSON, containing the tool name and parameters), which an external program executes and then returns the result, forming a "think—call—observe" loop; it is the engineering implementation of the ReAct paradigm.
At the end of the video, it is mentioned that the next episode will teach the model to call tools, indicating that the current deployment is only a starting point and that only local models plus tool-calling capabilities can cover more real-world workflows.
The model itself does not execute tools; instead, an external system (the Agent framework) parses the tool call request and actually runs the tool, then appends the result to the conversation history as text.
The accuracy of tool calling depends on the model's understanding of the task and the clarity of the tool descriptions.
Processing 10 invoices may trigger dozens of tool calls and introduce thousands of lines of irrelevant data.
There is a severe disconnect between models' verbal credibility ratings of fake notices and their actual actions: models still issued harmful tool calls in 27% to 99% of cases even when they had themselves rated the notice as untrustworthy.
Any model with tool-calling capability needs finer-grained sandbox isolation and network egress control.
In security practice, tool-calling permissions must be intercepted and authorized by the host application (such as an Agent framework) outside the model; the model's own output is only a "suggestion," not a "command."
40 more timeline events
All Facts (20)
Function Calling (Tool Use) is a standardized mechanism for large models to invoke external tools, where developers pre-declare tool names, descriptions, and parameter schemas, and the model autonomously determines when to make a call.
90%VerifiedThe underlying mechanism of Tool Use is that developers pre-define the name, description, and parameter structure (Schema) of tools, and the model determines during inference when and which tool to call and outputs structured invocation instructions.
90%VerifiedThe tool-calling (Function Calling) mechanism allows the model to invoke external tools during reasoning, such as file read/write operations, terminal commands, and search APIs.
90%VerifiedAI Agent通过工具调用（Tool Use）机制与外部系统交互，包括执行代码、调用API、读写数据库等
90%VerifiedThe core mechanism of Function Calling is that developers describe external function signatures in JSON Schema format, and the model outputs structured JSON invocation instructions that conform to the schema specification.
80%VerifiedHarness Engineering概念源自软件测试领域的Test Harness（测试框架），指围绕被测系统搭建的一套自动化测试基础设施
80%VerifiedFunction Calling机制中，开发者将工具的名称、功能描述、参数Schema（通常为JSON Schema格式）注入到模型的请求中，模型输出结构化的JSON而非自然语言来调用工具
80%VerifiedAI Agent与传统聊天AI的根本区别在于具备感知-决策-行动的闭环能力，采用ReAct（Reasoning + Acting）框架
80%VerifiedAgents use the Tool Calling mechanism to allow models to output calling intent in a structured JSON format, with external programs executing the actual operations and feeding results back to the model.
80%VerifiedIn the ReAct framework, when the number of tools exceeds 20, both token consumption and the model's tool selection accuracy drop significantly, a problem known as 'tool overload'.
80%VerifiedAgent Skills是Anthropic发布的能力封装标准，定义为一个包含Skill.md文件的目录，可包含说明、脚本和其他资源
80%VerifiedFunction 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.
80%VerifiedThe 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.
75%VerifiedA managed Agent API means the platform is responsible for runtime management of agents, including session state maintenance, tool call orchestration, error retries, monitoring, logging, and other infrastructure work
75%VerifiedEvaluation 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
75%VerifiedAgents (intelligent agents) can invoke tools on a computer and independently complete an entire set of complex tasks, unlike chatbots that can only conduct one-round-at-a-time conversations.
75%VerifiedThe 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.
75%VerifiedThe basic principle of tool calling is that the model identifies scenarios requiring external capabilities, outputs a structured function call request (usually in JSON format), and after the external system executes it, the result is returned to the model to continue reasoning.
75%VerifiedThe key breakthrough in tool-calling capability lies in the model's ability to generate structured function-call requests and correctly parse tool return results before continuing to reason, rather than simply concatenating tool outputs into the prompt
75%VerifiedTool calling allows models to dynamically invoke external functions during inference and incorporate execution results into the context of subsequent reasoning, forming a closed loop of perception, decision-making, and execution.
70%