Skip to content

Protocols API

SubAgentDepsProtocol

subagents_pydantic_ai.SubAgentDepsProtocol

Bases: Protocol

Protocol for dependencies that support subagent management.

The only method the library calls is clone_for_subagent, which decides what a delegated subagent inherits from its parent. Your deps class may be frozen or use slots -- the library never writes attributes onto it.

Example
Python
@dataclass(slots=True)
class MyDeps:
    backend: Backend

    def clone_for_subagent(self, max_depth: int = 0) -> "MyDeps":
        return MyDeps(backend=self.backend)

Earlier versions also required a subagents: dict[str, Any] attribute. The library never read it, so every application carried a field for nothing; the requirement is gone. A deps class that still declares it satisfies this protocol unchanged.

Source code in src/subagents_pydantic_ai/protocols.py
Python
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
@runtime_checkable
class SubAgentDepsProtocol(Protocol):
    """Protocol for dependencies that support subagent management.

    The only method the library calls is `clone_for_subagent`, which decides what
    a delegated subagent inherits from its parent. Your deps class may be frozen
    or use `slots` -- the library never writes attributes onto it.

    Example:
        ```python
        @dataclass(slots=True)
        class MyDeps:
            backend: Backend

            def clone_for_subagent(self, max_depth: int = 0) -> "MyDeps":
                return MyDeps(backend=self.backend)
        ```

    Earlier versions also required a `subagents: dict[str, Any]` attribute. The
    library never read it, so every application carried a field for nothing; the
    requirement is gone. A deps class that still declares it satisfies this
    protocol unchanged.
    """

    def clone_for_subagent(self, max_depth: int = 0) -> SubAgentDepsProtocol:
        """Create a new deps instance for a delegated subagent.

        Return a **new** instance rather than `self`. The library relies on each
        delegation getting its own deps, and a shared instance leaks state
        between concurrent subagents.

        A subagent typically inherits shared resources (a backend, a filesystem)
        and gets fresh task-specific state.

        Args:
            max_depth: Remaining nesting budget, one less than the parent's
                `max_nesting_depth`. At or below zero the subagent should not be
                given the means to delegate further. Enforcement is yours: the
                library passes the number through but does not itself refuse a
                nested delegation.

        Returns:
            A new deps instance configured for the subagent.
        """
        ...

clone_for_subagent(max_depth=0)

Create a new deps instance for a delegated subagent.

Return a new instance rather than self. The library relies on each delegation getting its own deps, and a shared instance leaks state between concurrent subagents.

A subagent typically inherits shared resources (a backend, a filesystem) and gets fresh task-specific state.

Parameters:

Name Type Description Default
max_depth int

Remaining nesting budget, one less than the parent's max_nesting_depth. At or below zero the subagent should not be given the means to delegate further. Enforcement is yours: the library passes the number through but does not itself refuse a nested delegation.

 0

Returns:

Type Description
SubAgentDepsProtocol

A new deps instance configured for the subagent.
Source code in src/subagents_pydantic_ai/protocols.py
Python
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
def clone_for_subagent(self, max_depth: int = 0) -> SubAgentDepsProtocol:
    """Create a new deps instance for a delegated subagent.

    Return a **new** instance rather than `self`. The library relies on each
    delegation getting its own deps, and a shared instance leaks state
    between concurrent subagents.

    A subagent typically inherits shared resources (a backend, a filesystem)
    and gets fresh task-specific state.

    Args:
        max_depth: Remaining nesting budget, one less than the parent's
            `max_nesting_depth`. At or below zero the subagent should not be
            given the means to delegate further. Enforcement is yours: the
            library passes the number through but does not itself refuse a
            nested delegation.

    Returns:
        A new deps instance configured for the subagent.
    """
    ...

MessageBusProtocol

subagents_pydantic_ai.MessageBusProtocol

Bases: Protocol

Protocol for message bus implementations.

The message bus enables communication between agents, supporting both fire-and-forget messages and request-response patterns.

Implementations can use different backends: - In-memory (default): Uses asyncio queues - Redis: For distributed multi-process setups - Custom: Any backend implementing this protocol

Example
Python
bus = InMemoryMessageBus()

# Register an agent
queue = bus.register_agent("worker-1")

# Send a message
await bus.send(AgentMessage(
    type=MessageType.TASK_UPDATE,
    sender="parent",
    receiver="worker-1",
    payload={"status": "starting"},
))

# Request-response pattern
response = await bus.ask(
    sender="parent",
    receiver="worker-1",
    question="What is your status?",
    task_id="task-123",
    timeout=30.0,
)
Source code in src/subagents_pydantic_ai/protocols.py
Python
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
@runtime_checkable
class MessageBusProtocol(Protocol):
    """Protocol for message bus implementations.

    The message bus enables communication between agents, supporting
    both fire-and-forget messages and request-response patterns.

    Implementations can use different backends:
    - In-memory (default): Uses asyncio queues
    - Redis: For distributed multi-process setups
    - Custom: Any backend implementing this protocol

    Example:
        ```python
        bus = InMemoryMessageBus()

        # Register an agent
        queue = bus.register_agent("worker-1")

        # Send a message
        await bus.send(AgentMessage(
            type=MessageType.TASK_UPDATE,
            sender="parent",
            receiver="worker-1",
            payload={"status": "starting"},
        ))

        # Request-response pattern
        response = await bus.ask(
            sender="parent",
            receiver="worker-1",
            question="What is your status?",
            task_id="task-123",
            timeout=30.0,
        )
        ```
    """

    async def send(self, message: AgentMessage) -> None:
        """Send a message to a specific agent.

        Args:
            message: The message to send. Must have a valid receiver.

        Raises:
            KeyError: If the receiver is not registered.
        """
        ...

    async def ask(
        self,
        sender: str,
        receiver: str,
        question: Any,
        task_id: str,
        timeout: float = 30.0,
    ) -> AgentMessage:
        """Send a question and wait for a response.

        This implements a request-response pattern where the sender
        blocks until the receiver answers or the timeout expires.

        Args:
            sender: ID of the asking agent.
            receiver: ID of the agent to ask.
            question: The question payload.
            task_id: Task ID for correlation.
            timeout: Maximum time to wait for response in seconds.

        Returns:
            The response message from the receiver.

        Raises:
            asyncio.TimeoutError: If no response within timeout.
            KeyError: If the receiver is not registered.
        """
        ...

    async def answer(self, original: AgentMessage, answer: Any) -> None:
        """Answer a previously received question.

        Args:
            original: The original question message.
            answer: The answer payload.
        """
        ...

    def register_agent(self, agent_id: str) -> asyncio.Queue[AgentMessage]:
        """Register an agent to receive messages.

        Args:
            agent_id: Unique identifier for the agent.

        Returns:
            A queue where messages for this agent will be delivered.

        Raises:
            ValueError: If agent_id is already registered.
        """
        ...

    def unregister_agent(self, agent_id: str) -> None:
        """Unregister an agent from the message bus.

        After unregistration, messages sent to this agent will raise errors.

        Args:
            agent_id: The agent to unregister.
        """
        ...

    async def get_messages(
        self,
        agent_id: str,
        timeout: float = 0.0,
    ) -> list[AgentMessage]:
        """Get pending messages for an agent.

        Non-blocking retrieval of all pending messages in the agent's queue.
        Optionally waits up to `timeout` seconds for at least one message.

        Args:
            agent_id: The agent to get messages for.
            timeout: Maximum time to wait for a message (0 = no wait).

        Returns:
            List of pending messages (may be empty).

        Raises:
            KeyError: If the agent is not registered.
        """
        ...

send(message) async

Send a message to a specific agent.

Parameters:

Name Type Description Default
message AgentMessage

The message to send. Must have a valid receiver.

 required

Raises:

Type Description
KeyError

If the receiver is not registered.
Source code in src/subagents_pydantic_ai/protocols.py
Python
101
102
103
104
105
106
107
108
109
110
async def send(self, message: AgentMessage) -> None:
    """Send a message to a specific agent.

    Args:
        message: The message to send. Must have a valid receiver.

    Raises:
        KeyError: If the receiver is not registered.
    """
    ...

ask(sender, receiver, question, task_id, timeout=30.0) async

Send a question and wait for a response.

This implements a request-response pattern where the sender blocks until the receiver answers or the timeout expires.

Parameters:

Name Type Description Default
sender str

ID of the asking agent.

 required
receiver str

ID of the agent to ask.

 required
question Any

The question payload.

 required
task_id str

Task ID for correlation.

 required
timeout float

Maximum time to wait for response in seconds.

 30.0

Returns:

Type Description
AgentMessage

The response message from the receiver.

Raises:

Type Description
TimeoutError

If no response within timeout.
KeyError

If the receiver is not registered.
Source code in src/subagents_pydantic_ai/protocols.py
Python
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
async def ask(
    self,
    sender: str,
    receiver: str,
    question: Any,
    task_id: str,
    timeout: float = 30.0,
) -> AgentMessage:
    """Send a question and wait for a response.

    This implements a request-response pattern where the sender
    blocks until the receiver answers or the timeout expires.

    Args:
        sender: ID of the asking agent.
        receiver: ID of the agent to ask.
        question: The question payload.
        task_id: Task ID for correlation.
        timeout: Maximum time to wait for response in seconds.

    Returns:
        The response message from the receiver.

    Raises:
        asyncio.TimeoutError: If no response within timeout.
        KeyError: If the receiver is not registered.
    """
    ...

answer(original, answer) async

Answer a previously received question.

Parameters:

Name Type Description Default
original AgentMessage

The original question message.

 required
answer Any

The answer payload.

 required
Source code in src/subagents_pydantic_ai/protocols.py
Python
141
142
143
144
145
146
147
148
async def answer(self, original: AgentMessage, answer: Any) -> None:
    """Answer a previously received question.

    Args:
        original: The original question message.
        answer: The answer payload.
    """
    ...

register_agent(agent_id)

Register an agent to receive messages.

Parameters:

Name Type Description Default
agent_id str

Unique identifier for the agent.

 required

Returns:

Type Description
Queue[AgentMessage]

A queue where messages for this agent will be delivered.

Raises:

Type Description
ValueError

If agent_id is already registered.
Source code in src/subagents_pydantic_ai/protocols.py
Python
150
151
152
153
154
155
156
157
158
159
160
161
162
def register_agent(self, agent_id: str) -> asyncio.Queue[AgentMessage]:
    """Register an agent to receive messages.

    Args:
        agent_id: Unique identifier for the agent.

    Returns:
        A queue where messages for this agent will be delivered.

    Raises:
        ValueError: If agent_id is already registered.
    """
    ...

unregister_agent(agent_id)

Unregister an agent from the message bus.

After unregistration, messages sent to this agent will raise errors.

Parameters:

Name Type Description Default
agent_id str

The agent to unregister.

 required
Source code in src/subagents_pydantic_ai/protocols.py
Python
164
165
166
167
168
169
170
171
172
def unregister_agent(self, agent_id: str) -> None:
    """Unregister an agent from the message bus.

    After unregistration, messages sent to this agent will raise errors.

    Args:
        agent_id: The agent to unregister.
    """
    ...

get_messages(agent_id, timeout=0.0) async

Get pending messages for an agent.

Non-blocking retrieval of all pending messages in the agent's queue. Optionally waits up to timeout seconds for at least one message.

Parameters:

Name Type Description Default
agent_id str

The agent to get messages for.

 required
timeout float

Maximum time to wait for a message (0 = no wait).

 0.0

Returns:

Type Description
list[AgentMessage]

List of pending messages (may be empty).

Raises:

Type Description
KeyError

If the agent is not registered.
Source code in src/subagents_pydantic_ai/protocols.py
Python
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
async def get_messages(
    self,
    agent_id: str,
    timeout: float = 0.0,
) -> list[AgentMessage]:
    """Get pending messages for an agent.

    Non-blocking retrieval of all pending messages in the agent's queue.
    Optionally waits up to `timeout` seconds for at least one message.

    Args:
        agent_id: The agent to get messages for.
        timeout: Maximum time to wait for a message (0 = no wait).

    Returns:
        List of pending messages (may be empty).

    Raises:
        KeyError: If the agent is not registered.
    """
    ...

The in-memory bus, TaskManager, and DynamicAgentRegistry are documented on their own pages: Message Bus and Registry.

Usage Examples

Implementing SubAgentDepsProtocol

Python
from dataclasses import dataclass, field
from typing import Any

@dataclass
class MyDeps:
    """Custom dependencies implementing SubAgentDepsProtocol."""

    subagents: dict[str, Any] = field(default_factory=dict)
    database_url: str = ""
    api_key: str = ""

    def clone_for_subagent(self, max_depth: int = 0) -> "MyDeps":
        """Create isolated deps for subagent."""
        return MyDeps(
            subagents={} if max_depth <= 0 else self.subagents.copy(),
            database_url=self.database_url,  # Share read-only config
            api_key=self.api_key,
        )

Implementing Custom Message Bus

Python
from subagents_pydantic_ai import MessageBusProtocol, AgentMessage

class RedisMessageBus:
    """Redis-based message bus for distributed systems."""

    def __init__(self, redis_url: str):
        self.redis = Redis.from_url(redis_url)

    async def send(self, message: AgentMessage) -> None:
        channel = f"agent:{message.receiver}"
        await self.redis.publish(channel, message.json())

    async def receive(
        self,
        agent_id: str,
        timeout: float | None = None,
    ) -> AgentMessage | None:
        # Implementation...
        pass

    async def subscribe(self, agent_id: str) -> None:
        pass

    async def unsubscribe(self, agent_id: str) -> None:
        pass

Using TaskManager

Python
from subagents_pydantic_ai import TaskManager, InMemoryMessageBus

bus = InMemoryMessageBus()
manager = TaskManager(message_bus=bus)

# Create a task
handle = await manager.create_task(
    subagent_name="researcher",
    description="Research Python async",
)

# Check status
status = await manager.get_task_status(handle.task_id)

# Answer a question
await manager.answer_question(handle.task_id, "Use asyncio")

# Cancel a task
await manager.cancel_task(handle.task_id, hard=False)

Using DynamicAgentRegistry

Python
from subagents_pydantic_ai import DynamicAgentRegistry

registry = DynamicAgentRegistry()

# List registered agents
agents = registry.list_agents()

# Get a specific agent
agent = registry.get_agent("custom-analyst")

# Remove an agent
registry.remove_agent("custom-analyst")