Tool
An external functional module that an AI agent can invoke to extend the boundaries of the agent's capabilities, enabling it to perform actual operations such as querying a database, sending email, or calling a sub-workflow.
Core Facts
Timeline (last 90 days)
LangChain abstracts the round-trip flow of tool calling into Tool objects and an agent execution loop, so that logic such as multi-step tool orchestration, result backfilling, and error retry need not be implemented from scratch.
Everything Claude does to a user's account is done through Tools; Claude will not bypass tools to directly manipulate a user's account, files or system.
The benefits brought by the tool middleware architecture include: controllability (precisely knowing what Claude is allowed to do), auditability (every tool call is recorded and can be checked), and security (operations without a defined tool cannot be executed).
When the model requests a tool call, the finish_reason field in the response has the value tool_calls rather than stop, which is the standard signal for determining whether the model is requesting a tool call.
Tool Use implementations from different model providers vary in parameter formats, parallel call support, streaming tool calls, and other aspects, and the AI SDK's unified abstraction layer is used to shield these differences.
All Facts (5)
Tool Use implementations from different model providers vary in parameter formats, parallel call support, streaming tool calls, and other aspects, and the AI SDK's unified abstraction layer is used to shield these differences.
75%UnverifiedLangChain abstracts the round-trip flow of tool calling into Tool objects and an agent execution loop, so that logic such as multi-step tool orchestration, result backfilling, and error retry need not be implemented from scratch.
50%UnverifiedEverything Claude does to a user's account is done through Tools; Claude will not bypass tools to directly manipulate a user's account, files or system.
50%UnverifiedThe benefits brought by the tool middleware architecture include: controllability (precisely knowing what Claude is allowed to do), auditability (every tool call is recorded and can be checked), and security (operations without a defined tool cannot be executed).
50%UnverifiedWhen the model requests a tool call, the finish_reason field in the response has the value tool_calls rather than stop, which is the standard signal for determining whether the model is requesting a tool call.
50%