flyte.ai.agents

flyte.ai.agents — Agent abstractions for Flyte apps.

Directory

Classes

Class	Description
`Agent`	A flyte-native tool-use agent harness.
`AgentEvent`	Lightweight event emitted by the agent loop.
`AgentResult`	Outcome of a single agent invocation.
`AgentTool`	A normalized tool descriptor used by :class:`Agent`.
`CodeModeAgent`	Generates code via an LLM, executes it in a Monty sandbox, and.
`LLMMessage`	Provider-agnostic shape returned by :data:`LLMCallable`.
`MCPServerSpec`	Declarative spec for a remote MCP server that exposes tools.
`MemoryMeta`	Per-file metadata sidecar (sha256, actor, timestamp, …) for a memory entry.
`MemoryStore`	Conversation transcript + path-addressed artifact memory backed by :class:`flyte.

Protocols

Protocol	Description
`AgentProtocol`	Minimal protocol that any agent must satisfy to work with.

Errors

Exception	Description
`AccessDenied`	Raised when a write targets a read-only or reserved prefix.
`ConcurrencyError`	Raised when an `expected_sha` precondition does not match the current state.
`MemoryStoreError`	Base class for :class:`MemoryStore` errors.

Methods

Method	Description
`tool()`	Wrap a task, ``@flyte.

Variables

Property	Type	Description
`agent_progress_cb`	`ContextVar`

Methods

tool()

        
    
def tool(
    obj: Any,
    name: str | None,
    description: str | None,
    requires_approval: bool,
) -> AgentTool | Callable[[Any], AgentTool]

Wrap a task, @flyte.trace helper, plain callable, or LazyEntity as an :class:AgentTool.

This removes the boilerplate of building an :class:AgentTool by hand (manually pulling the docstring, JSON schema, and writing a dict-args execution bridge) when you only need to tweak how a tool is presented to the model or gate it behind human approval.

Use it as a direct call::

refund_tool = tool(issue_refund, requires_approval=True)

or as a (parametrized) decorator stacked on top of @env.task / @flyte.trace / a plain function::

@tool(requires_approval=True)
@env.task
async def issue_refund(order_id: str, amount_usd: float) -&gt; dict: ...

@tool
def search(query: str) -&gt; str: ...

The wrapped task is still registered with its :class:~flyte.TaskEnvironment and executes on-cluster via task.aio when the agent calls it.

Parameter	Type	Description
`obj`	`Any`	The object to wrap. Omit when using `tool` as a parametrized decorator (`@tool(...)`).
`name`	`str \| None`	Override the tool name shown to the model. Defaults to the function / task name.
`description`	`str \| None`	Override the description shown to the model. Defaults to the first paragraph of the object’s docstring.
`requires_approval`	`bool`	Gate execution behind the agent’s HITL approval callback.

Returns: An :class:AgentTool (direct call) or a decorator returning one.