Agent Tools

Convert an object, such as an R function or an agent, into a Tool object for use by an agent. The equip function performs this conversion automatically, but by calling tool directly, the user can customize the tool before equipping it.

tool(x, name = deparse(substitute(x)), ...)

Arguments

x

Object to convert to a tool, such as a function or Agent

name

Name of the new tool

...

Arguments passed to convert. Methods on convert typically forward these to the Tool constructor.

Value

A Tool object, which is an R function decorated with metadata, such as the argument types and return type.

Details

The more annotations on a tool, the more the agent will understand how to use it. describe_with_Rd is a convenient way to annotate a tool based on a packaged R function with information from its man page.

Arguments passed to a tool are converted according to the types declared by can_accept_as, without any conversion by default.

Converting an Agent to a Tool enables an agent to be called by another agent (enabling networks of agents), as well as the user by simply invoking it like any other function, fully realizing the goal of the wizrd package: to program agents as functions.

The call to the tool is evaluated in the env provided to chat, so any arguments converted to language objects will be evaluated as R code. If the agent sends a backquoted string, like `x`, it will be converted to a name and resolved dynamically, as shown in the examples.

There is nothing LLM-specific about Tool objects. Thus, they could even be used as a way to annotate ordinary R functions, like with inferred parameter types.

Author

Michael Lawrence

See also

equip for equipping an agent with a tool; functions like can_accept_as, demonstrate and describe_with_Rd for further specifying a tool

Examples

if (FALSE) { # \dontrun{
    var <- 1:10
    equip(llama(), tool(mean) |> describe_with_Rd()) |>
        predict("What is the mean of `var`?")
} # }