--- title: Python description: Build durable workflows and AI agents in Python with the Vercel SDK. type: guide summary: Set up the Workflow Python SDK in your Python application. prerequisites: - /docs/getting-started related: - /docs/foundations - /docs/foundations/workflows-and-steps --- # Python The Python SDK is currently in **beta**. APIs and behavior may change. You can build durable workflows in Python using the [`vercel-workflow` SDK](https://pypi.org/project/vercel-workflow/). Your workflow code can pause, resume, and maintain state, just like the JavaScript and TypeScript Workflow SDK. ## Getting started Add the `vercel-workflow` package and workflow entrypoint to `pyproject.toml`: ```toml filename="pyproject.toml" [project] requires-python = ">=3.12" dependencies = ["vercel-workflow"] [[tool.vercel.workflows]] entrypoint = "app.workflows:wf" ``` The workflow `entrypoint` uses the `module:object` format and points to the exported `Workflows` registry. ## Workflows A workflow is a stateful function that coordinates multi-step logic over time. Create a `Workflows` instance and use the `@wf.workflow` decorator to mark a function as durable: ```python filename="app/workflow.py" from vercel import workflow wf = workflow.Workflows( sandbox_policy=workflow.SandboxPolicy(share_sandboxes=True), # [!code highlight] ) ``` `share_sandboxes=True` reuses a sandbox across workflow runs for faster startup. Module globals are shared between those runs, so workflow code should not mutate global state. ```python filename="app/workflows/ai_content_workflow.py" from app.workflow import wf from app.steps.generate_draft import generate_draft, summarize_draft @wf.workflow # [!code highlight] async def ai_content_workflow(*, topic: str): draft = await generate_draft(topic=topic) summary = await summarize_draft(draft=draft) return { "draft": draft, "summary": summary, } ``` Export the registry from the workflow package and import the module containing your workflow so its definitions are registered: ```python filename="app/workflows/__init__.py" from app.workflow import wf from app.workflows import ai_content_workflow __all__ = ["ai_content_workflow", "wf"] ``` Under the hood, the workflow compiles into a route that orchestrates execution. All inputs and outputs are recorded in an event log. If a deploy or crash happens, the system replays execution deterministically from where it stopped. ## Steps A step is a stateless function that runs a unit of durable work inside a workflow. Use `@wf.step` to mark a function as a step: ```python filename="app/steps/generate_draft.py" import random from app.workflow import wf @wf.step # [!code highlight] async def generate_draft(*, topic: str): return await ai_generate(prompt=f"Write a blog post about {topic}") @wf.step # [!code highlight] async def summarize_draft(*, draft: str): summary = await ai_summarize(text=draft) # Simulate a transient error. The step automatically retries. if random.random() < 0.3: raise Exception("Transient AI provider error") return summary ``` Each step executes separately from the workflow orchestrator. While the step executes, the workflow suspends without consuming resources. When the step completes, the workflow resumes automatically where it left off. ### Cancellable steps Pass `cancellable=True` when a workflow may need to stop a running step. This makes normal Python `asyncio` cancellation apply to the step: if a task awaiting the step is cancelled in the workflow, then the task running the step will be cancelled as well. For example, `asyncio.timeout()` cancels the step if it does not finish within the allotted time: ```python filename="app/workflows/report.py" import asyncio from datetime import timedelta from app.workflow import wf @wf.step(cancellable=True) # [!code highlight] async def generate_report(*, account_id: str) -> dict[str, str]: return await reporting_service.generate(account_id=account_id) @wf.workflow async def report_workflow(*, account_id: str) -> dict[str, str]: try: async with asyncio.timeout(timedelta(minutes=10).total_seconds()): # [!code highlight] return await generate_report(account_id=account_id) except TimeoutError: return {"status": "timed_out"} ``` Cancellation is a request, so the workflow waits for the step to terminate before continuing. If the step suppresses its cancellation, it can still return normally. ## Starting a workflow Call `workflow.start()` from server-side code to start a workflow. It returns a `Run` that you can use to identify the run, check its status, and wait for its result: ```python filename="app/api/generate.py" from app.workflows.ai_content_workflow import ai_content_workflow from vercel import workflow @app.post("/api/generate") async def generate_content(*, topic: str): run = await workflow.start(ai_content_workflow, topic=topic) # [!code highlight] print(run.run_id) print(await run.status()) # [!code highlight] # Wait until the workflow completes and return its result. return await run.return_value() # [!code highlight] ``` Starting a workflow only waits until the run has been created and queued. Await `return_value()` to wait for the workflow to finish, or save its `run_id` and recreate the handle later with `workflow.Run(run_id)`. ## Sleep Sleep pauses a workflow for a specified duration without consuming compute resources: ```python filename="app/workflows/ai_refine.py" from datetime import timedelta from app.workflow import wf from vercel import workflow @wf.workflow async def ai_refine_workflow(*, draft_id: str): draft = await fetch_draft(draft_id) await workflow.sleep(timedelta(days=7)) # Wait 7 days to gather more signals. # [!code highlight] refined = await refine_draft(draft) return { "draft_id": draft_id, "refined": refined, } ``` The parameter accepts four forms: | Form | Description | Example | | -------------------- | ---------------------------------------------- | ---------------------------------- | | `str` | Human-readable duration string | `"2 days"`, `"1w"`, `"1h 30m"` | | `int` or `float` | Seconds from now | `5` (5 seconds) | | `datetime.timedelta` | Duration from now | `timedelta(days=7)` | | `datetime.datetime` | Absolute wake-up time (must be timezone-aware) | `datetime(2025, 1, 1, tzinfo=UTC)` | The string form accepts one or more `` pairs. Supported units: | Duration | Unit | | ------------ | ------------------------ | | Milliseconds | `ms` | | Seconds | `s`, `second`, `seconds` | | Minutes | `m`, `minute`, `minutes` | | Hours | `h`, `hour`, `hours` | | Days | `d`, `day`, `days` | | Weeks | `w`, `week`, `weeks` | `sleep()` must be called from the workflow body, not from inside a step. Calling it from a step raises a `RuntimeError`. The sleep consumes no resources. The workflow resumes automatically when the time expires. `asyncio.sleep()` may also be used to sleep. ## Deterministic workflow helpers Workflow bodies must be fully deterministic, and so are not allowed to perform operations like reading the system clock or generating randomness using the default generator. Deterministic replacements are provided for use in workflow bodies: * `workflow.now()` is a deterministic substitute for `datetime.datetime.now`. It returns the time that the last workflow event occurred at. * `workflow.time_ns()` is a deterministic substitute for `time.time_ns`. * `workflow.random()` returns a `random.Random` instance with a seed based on the run id. These helpers can only be called from a workflow body. Steps should use the normal system functions. ## Hooks A hook lets a workflow wait for external events such as user actions, webhooks, or third-party API responses. Define a hook model with Pydantic and `workflow.BaseHook`: ```python filename="app/workflows/approval.py" import typing import pydantic from app.workflow import wf from vercel import workflow class Approval(pydantic.BaseModel, workflow.BaseHook): # [!code highlight] """Human approval for AI-generated drafts""" decision: typing.Literal["approved", "changes"] notes: str | None = None @wf.workflow async def ai_approval_workflow(*, topic: str) -> None: draft = await generate_draft(topic=topic) # Wait for human approval events async for event in Approval.wait(token="draft-123"): # [!code highlight] if event.decision == "approved": await publish_draft(draft) break revised = await refine_draft(draft, event.notes) await publish_draft(revised) ``` Resume the workflow when data arrives: ```python filename="app/api/resume.py" from app.workflows.approval import Approval @app.post("/api/resume") async def resume(approval: Approval): # [!code highlight] """Resume the workflow when an approval is received""" hook = await approval.resume("draft-123") # [!code highlight] print(f"Resumed workflow run: {hook.run_id}") # [!code highlight] return {"ok": True} ``` When a hook receives data, the workflow resumes automatically. You don't need polling, message queues, or manual state management. ## Streaming Steps can stream progress while a workflow is running. Streams are associated with workflow runs, can be acquired from inside a workflow or inside a step, and can be passed to workflows and steps as arguments. If a type is specified when getting the stream (or on the type annotation of a step or workflow it is passed to), then Pydantic will be used to encode and decode the type. ```python filename="app/workflows/streaming.py" import pydantic from app.workflow import wf from vercel import workflow class Progress(pydantic.BaseModel): message: str @wf.step async def write_progress( writable: workflow.WorkflowWritable[Progress], # [!code highlight] ): for message in ["Drafting", "Reviewing", "Complete"]: await writable.write(Progress(message=message)) # [!code highlight] await writable.close() @wf.workflow async def streaming_workflow(): writable = workflow.get_writable(type=Progress) await write_progress(writable) # [!code highlight] ``` Read and validate the typed values from the returned `Run` as they arrive: ```python filename="app/api/stream.py" from app.workflows.streaming import Progress, streaming_workflow from vercel import workflow @app.post("/api/stream") async def stream_progress(): run = await workflow.start(streaming_workflow) async for progress in run.readable(type=Progress): # [!code highlight] print(progress.message) ``` Streams are not closed automatically. It can be manually closed when it is finished so that readers know to terminate. `readable` and `get_writable` also take a `namespace` parameter, allowing each run to have many distinct streams. ## Next steps * Learn more about the [Foundations](/docs/foundations). * Check [Errors](/docs/errors) if you encounter issues. * Explore the [API Reference](/docs/api-reference). --- For a semantic overview of all documentation, see [/sitemap.md](/sitemap.md) For an index of all available documentation, see [/llms.txt](/llms.txt) For agent-facing discovery, including API and MCP surfaces, see [/agents.md](/agents.md)