Structured Tools

This post from the LangChain team announces Structured Tools in LangChain: a new tool abstraction that lets a tool take an arbitrary number of inputs of arbitrary types, where earlier LangChain tools took a single string. It pairs the abstraction with a new agent class, StructuredChatAgent, designed to work with such tools.

The post places the change in the history of LangChain’s agents. Tool use in the early days was simplistic — a model generated a tool name and one input string, which confined an agent to one tool per turn — and the post attributes that restriction mainly to what models of the time could reliably do. It credits more capable models such as text-davinci-003, gpt-3.5-turbo and gpt-4 with raising that floor, which prompted the team to revisit the limits on tool usage.

Key Points

• When LangChain launched in November 2022, agent and tool use were central to its design, and the team built one of the first chains based on the ReAct paper.
• Earlier in 2023 LangChain had introduced a “multi-action” agent framework in which an agent can plan several actions per step; structured tools remove the single-string input constraint on top of that.
• A structured tool wraps any function so that an agent can call it, and is defined by a name, a description, an args_schema (a Pydantic BaseModel that both tells the agent what inputs are required and validates them before execution) and _run/_arun functions holding the tool’s logic.
• The post advises that a tool’s name should communicate unambiguously what it does — “if a tool’s name isn’t clear to you, it probably isn’t clear to the agent either” — and that names can also hint at how tools relate to one another.
• StructuredTool.from_function() builds a tool from a callable and infers its args_schema from the function’s signature; subclassing BaseTool gives more control over the definition.
• Two new toolkits were released on the new class: file management (write, grep, move, copy, list directories, find) and a stateful PlayWright browser toolkit for visiting websites, clicking, submitting forms and querying data.
• Structured tools with more than one argument do not work with LangChain’s earlier agent types without customization, whereas tools that take a single string still work with existing agents, and older string tools work with the new agent.

Context

The post is the vendor’s own feature announcement, and its account of why earlier tools were limited to a single string — the constraints of the models of the time — is the team’s own explanation.