Existing Agents

An agent built on something other than AgentDeck or the OpenAI Agents SDK - a hand-rolled loop, LangGraph, anything with a Python entry point - runs under a Deck by calling it from a @workflow, unchanged:

from agentdeck import Deck, WorkflowCtx, workflow


class LegacyAgent:
    async def answer(self, question: str) -> str:
        ...  # whatever this already does


legacy = LegacyAgent()


@workflow
async def support(ctx: WorkflowCtx, question: str) -> str:
    return await legacy.answer(question)


deck = Deck(workflows=[support])

@workflow puts a name and a context contract around ordinary Python; the body is free to call anything, including an agent implementation that predates AgentDeck entirely. Nothing about legacy changes - it keeps whatever model client, prompt loop or framework it already uses.

Running it through deck.runs.start("support", "where is my order?") is what the wrap buys: this call gets a durable Run (pause/resume/cancel, rehydration from another process) and an event log, the same as an Agent or a native @tool would. ctx.ask(...) inside the body suspends the run for a human answer, if legacy ever needs one - see Workflows for what a workflow body can do beyond this one call.

Already have an agents.Agent built with the OpenAI Agents SDK directly? See OpenAI Agents SDK - it attaches without a @workflow wrapper.