AgentTool

Package: flyte.ai.agents

A normalized tool descriptor used by flyte.ai.agents.Agent.

Most users do not construct flyte.ai.agents.AgentTool directly — pass plain callables, @flyte.trace helpers, or @env.task templates to flyte.ai.agents.Agent and they will be wrapped automatically. Build one explicitly when you need to:

• rename a tool for the LLM,
• override the description shown to the model,
• require human approval before execution (HITL),
• inject a fully custom JSON schema,
• intercept invocation with a call_handler.

Parameters

class AgentTool(
    name: str,
    description: str,
    parameters: dict[str, Any],
    execute: _ToolExecutor,
    requires_approval: bool = False,
    source: Literal['function', 'task', 'trace', 'remote_task', 'mcp', 'custom'] = 'function',
    target: Any = None,
    call_handler: ToolCallHandler | None = None,
    call_llm: LLMCallable | None = None,
    model: str | None = None,
)

Methods

aio()

def aio(
    *args: Any,
    **kwargs: Any,
) -> Any

Invoke the tool, routing through call_handler when one is registered.

Mirrors flyte._task.TaskTemplate.aio enough for flyte.map and in-task calls on @tool-wrapped tasks. When a call_handler is set, it runs with AgentTool.call_llm and AgentTool.model (or their defaults). Otherwise, durable @env.task / remote-task targets delegate to their underlying .aio; everything else goes through AgentTool.execute.

to_openai_format()

def to_openai_format()

Convert to the OpenAI / litellm tools schema.