FastA2A A2A SDK

FastA2A

FastA2A is an agentic framework agnostic implementation of the A2A protocol in Python. The library is designed to be used with any agentic framework, and is not exclusive to Pydantic AI.

Installation

FastA2A is available on PyPI as fasta2a so installation is as simple as:

pip install fasta2a  # or `uv add fasta2a`

The only dependencies are:

• starlette: to expose the A2A server as an ASGI application
• pydantic: to validate the request/response messages
• opentelemetry-api: to provide tracing capabilities

Usage

To use FastA2A, you need to bring the Storage, Broker and Worker components.

FastA2A was designed with the mindset that the worker could, and should live outside the web server. i.e. you can have a worker that runs on a different machine, or even in a different process.

You can use the InMemoryStorage and InMemoryBroker to get started, but you'll need to implement the Worker to be able to execute the tasks with your agentic framework. Let's see an example:

import uuid
from collections.abc import AsyncIterator
from contextlib import asynccontextmanager
from typing import Any

from fasta2a import FastA2A, Worker
from fasta2a.broker import InMemoryBroker
from fasta2a.schema import Artifact, Message, Part, TaskIdParams, TaskSendParams
from fasta2a.storage import InMemoryStorage

Context = list[Message]
"""The shape of the context you store in the storage."""


class InMemoryWorker(Worker[Context]):
    async def run_task(self, params: TaskSendParams) -> None:
        task = await self.storage.load_task(params['id'])
        assert task is not None

        await self.storage.update_task(task['id'], state='working')

        context = await self.storage.load_context(task['context_id']) or []
        context.extend(task.get('history', []))

        # Call your agent here...
        message = Message(
            role='agent',
            parts=[Part(text=f'Your context is {len(context) + 1} messages long.')],
            message_id=str(uuid.uuid4()),
        )

        # Update the new message to the context.
        context.append(message)

        artifacts = self.build_artifacts(123)
        await self.storage.update_context(task['context_id'], context)
        await self.storage.update_task(task['id'], state='completed', new_messages=[message], new_artifacts=artifacts)

    async def cancel_task(self, params: TaskIdParams) -> None: ...

    def build_message_history(self, history: list[Message]) -> list[Any]: ...

    def build_artifacts(self, result: Any) -> list[Artifact]: ...


storage = InMemoryStorage()
broker = InMemoryBroker()
worker = InMemoryWorker(storage=storage, broker=broker)


@asynccontextmanager
async def lifespan(app: FastA2A) -> AsyncIterator[None]:
    async with app.task_manager:
        async with worker.run():
            yield


app = FastA2A(storage=storage, broker=broker, lifespan=lifespan)

You can run this example as is with uvicorn main:app --reload.

Using Pydantic AI

Initially, this FastA2A lived under Pydantic AI repository, but since we received community feedback, we've decided to move it to a separate repository.

[!NOTE] Other agentic frameworks are welcome to implement the Worker component, and we'll be happy add the reference here.

Install the integration with the pydantic-ai extra:

pip install 'fasta2a[pydantic-ai]'  # or `uv add 'fasta2a[pydantic-ai]'`

Then turn any pydantic_ai.Agent into an A2A-compatible ASGI app:

from pydantic_ai import Agent
from fasta2a.pydantic_ai import agent_to_a2a

agent = Agent('openai:gpt-5.5')
app = agent_to_a2a(agent)

You can run this example as is with uvicorn main:app --reload.

As you see, it's pretty easy from the point of view of the developer using your agentic framework.

[!NOTE] The bridge requires Pydantic AI 2. Agent.to_a2a() was removed there; agent_to_a2a is its replacement.

Streaming

message/stream answers with server-sent events: the task first, then whatever the Worker publishes while it runs — publish_status for a change of state, publish_artifact for a result, whole or chunk by chunk with append=True — and the stream ends when the task reaches a final state. The worker does not have to publish that end: once run_task or cancel_task returns, the task's state is read back from storage and, if the task is over (completed, canceled, failed, rejected) or waiting on the client (input-required, auth-required), published and the stream closed — failed when the run raised — so a worker that only writes storage still ends its stream. tasks/resubscribe on a task in one of those states returns the task and ends at once.

The Pydantic AI bridge publishes working when it starts, the model's text as chunks of the answer's artifact while it is written, and the whole artifact as the last chunk.

Extensions

An extension is a capability negotiated by URI on top of the core protocol. The agent declares the ones it supports in its card, a client asks for some of them in the A2A-Extensions request header, and the agent answers on the same header with the ones it activated.

Declare them on the application (or pass extensions= to agent_to_a2a):

from fasta2a import AgentExtension, FastA2A

app = FastA2A(
    storage=storage,
    broker=broker,
    extensions=[
        AgentExtension(uri='https://example.com/ext/citations/v1', description='Cites its sources'),
        AgentExtension(uri='https://example.com/ext/trace/v1', required=True),
    ],
)

For every request, FastA2A activates the requested extensions it supports — a URI it never declared is ignored and not echoed back, which is how the client learns that — and puts the activated list in the message metadata, so a Worker can read what was agreed for the task it is running:

from fasta2a import activated_extensions


class MyWorker(Worker[Context]):
    async def run_task(self, params: TaskSendParams) -> None:
        if 'https://example.com/ext/citations/v1' in activated_extensions(params):
            ...  # add citation parts to the artifacts

A message/send or message/stream that leaves a required extension inactive is refused with a -32600 error whose data.missing_required_extensions names it, before any task is created.

Output modes

The card says which media types the agent answers in: default_output_modes (and default_input_modes) on the application, application/json when unsaid, and each Skill's own output_modes. A mode that is not a media type (type/subtype) is refused when the application is made.

app = FastA2A(
    storage=storage,
    broker=broker,
    default_input_modes=['text/plain'],
    default_output_modes=['text/markdown', 'application/x-ipynb+json'],
)

A client names the media types it accepts for a request in its configuration (acceptedOutputModes), and the worker reads them from its task's params, accepted_output_modes, absent when the client named none. What the worker gives in those modes goes in its artifacts, each part carrying its media_type:

class MyWorker(Worker[Context]):
    async def run_task(self, params: TaskSendParams) -> None:
        if 'application/x-ipynb+json' in params.get('accepted_output_modes', []):
            ...  # add an artifact with a part of that media type

Design

FastA2A is built on top of Starlette, which means it's fully compatible with any ASGI server.

Given the nature of the A2A protocol, it's important to understand the design before using it, as a developer you'll need to provide some components:

• Storage: to save and load tasks and the conversation context
• Broker: to schedule tasks
• Worker: to execute tasks

Let's have a look at how those components fit together:

flowchart TB
    Server["HTTP Server"] <--> |Sends Requests/<br>Receives Results| TM

    subgraph CC[Core Components]
        direction RL
        TM["TaskManager<br>(coordinates)"] --> |Schedules Tasks| Broker
        TM <--> Storage
        Broker["Broker<br>(queues & schedules)"] <--> Storage["Storage<br>(persistence)"]
        Broker --> |Delegates Execution| Worker
    end

    Worker["Worker<br>(implementation)"]

FastA2A allows you to bring your own Storage, Broker and Worker.

You can also leverage the in-memory implementations of Storage and Broker by using the InMemoryStorage and InMemoryBroker:

from fasta2a import InMemoryStorage, InMemoryBroker

storage = InMemoryStorage()
broker = InMemoryBroker()

Tasks and Context

FastA2A is designed to be opinionated regarding the A2A protocol. When the server receives a message, according to the specification, the server can decide between:

• Send a stateless message back to the client
• Create a stateful Task and run it on the background

FastA2A will always create a Task and run it on the background (on the Worker).

[!NOTE] You can read more about it here.

• Task: Represents one complete execution of an agent. When a client sends a message to the agent, a new task is created. The agent runs until completion (or failure), and this entire execution is considered one task. The final output should be stored as a task artifact.

• Context: Represents a conversation thread that can span multiple tasks. The A2A protocol uses a context_id to maintain conversation continuity:

  • When a new message is sent without a context_id, the server generates a new one
  • Subsequent messages can include the same context_id to continue the conversation
  • All tasks sharing the same context_id have access to the complete message history

Storage

The Storage component serves two purposes:

1. Task Storage: Stores tasks in A2A protocol format, including their status, artifacts, and message history
2. Context Storage: Stores conversation context in a format optimized for the specific agent implementation

This design allows for agents to store rich internal state (e.g., tool calls, reasoning traces) as well as store task-specific A2A-formatted messages and artifacts.

License

This project is licensed under the MIT License - see the LICENSE file for details.