Integrating External Agents via A2A

New Beta Feature in 3.14Rasa supports stateful execution of external agents via A2A protocol.This feature is in a beta (experimental) stage and may change in future Rasa versions. We welcome your feedback on this feature.

Overview

Rasa can act as an intelligent orchestrator that coordinates a network of multiple agents through the Agent-to-Agent (A2A) protocol. This is particularly useful when at least one of the agents is externally built, i.e. by a different department, by a third party, or it’s a legacy project on a different tech stack. Integrating each of these agents into a single system enables contextually rich conversations through single unified conversational interface.

Inverse role — Rasa as an A2A sub-agentThis guide covers Rasa as the orchestrator that calls external A2A agents. If you need the opposite — an external orchestrator calling your Rasa assistant — see Exposing Rasa as an A2A Sub-Agent and the A2A Server reference.

The Orchestrator Model

In this architecture, your Rasa agent serves as the orchestrator that:

Owns the conversation flow - Determines when and which external agents to involve
Manages business logic - Enforces business rules, compliance, and conversation patterns
Controls agent handoffs - Decides when to delegate tasks and when to resume control
Maintains conversation context - Preserves user context across different agent interactions
Allows context engineering - Allows fine-grained control over context passed between agents
Handles final responses - Ensures consistent user experience regardless of which agent provided the information

Configuring External Sub-Agents

Registering the sub agent

Configure external sub agents in the sub_agents/ folder of your Rasa agent:

your_project/
├── config.yml
├── domain/
├── data/flows/
└── sub_agents/
    ├── analytics_agent/
       └── config.yml

The A2A protocol prescribes using an agent card to expose any externally running agent’s capabilities. The agent card can be provided as part of the config.yml defined for each external agent -

sub_agents/analytics_agent/config.yml

# Basic agent information
agent:
  name: analytics_agent
  protocol: a2a
  description: "External agent that provides analytical capabilities"

# A2A configuration
configuration:
  agent_card: ./sub_agents/shopping_agent/agent_card.json

More configuration options can be found in the reference page.

Supporting long-running tasks

If the external A2A agent performs long-running work, configure background push mode so Rasa can keep the conversation responsive while the remote task continues. The external agent card must advertise capabilities.pushNotifications: true, and the sub-agent configuration must include the public Rasa base URL: The agent_card value can be a local file path or a URL.

sub_agents/analytics_agent/config.yml

agent:
  name: analytics_agent
  protocol: A2A
  description: "External agent that provides analytical capabilities"

configuration:
  agent_card: ./sub_agents/analytics_agent/agent_card.json
  push_notification_url: https://rasa.example.com

Rasa appends /agent/task_update to push_notification_url. When a background A2A task is still running and the user sends another message, Rasa triggers pattern_external_agent_processing; the default response is utter_agent_busy. For channel-specific delivery behavior and fallback behavior, see Long-running tasks with push notifications.

Invoking the External Agent

The orchestrating layer in Rasa provides you the control to invoke an external agent when your business logic needs you to do so by using the call step in a flow -

flows.yml

flows:
  product_enquiry:
    description: Handle questions related to analysis needed on the product
    steps:
    - collect: query
    - call: analytics_agent  # External A2A agent

For more information on how Rasa passes control and receives response from an external agent via A2A, head over to the documentation on External Sub Agents.

Context Management

The orchestrating agent in Rasa can control what context is shared with external agents by customizing the client interfacing with the external agent. To do so, define a custom python module -

sub_agents/analytics_agent/custom_agent.py

from rasa.agents.protocol.a2a.a2a_agent import A2AAgent
from rasa.agents.schemas import AgentInput

class AnalyticsAgent(A2AAgent):
    async def process_input(self, input: AgentInput) -> AgentInput:
        # Filter slots to only include relevant context
        filtered_slots = [
            slot for slot in input.slots
            if slot.name in ["user_id", "account_type", "inquiry_category"]
        ]

        # Create new input with filtered context
        return AgentInput(
            id=input.id,
            user_message=input.user_message,
            slots=filtered_slots,  # Only share necessary slots
            conversation_history=input.conversation_history,
            events=input.events,
            metadata=input.metadata,
            timestamp=input.timestamp
        )

Provide the customized interface to your sub agent’s config:

sub_agents/analytics_agent/config.yml

agent:
  name: analytics_agent
  protocol: a2a
  description: "External agent that provides analytical capabilities"

configuration:
  agent_card: ./sub_agents/analytics_agent/agent_card.json
  module: sub_agents.analytics_agent.custom_agent.AnalyticsAgent

For more information on possible customizations, head over to the reference documentation.

​Overview

​The Orchestrator Model

​Configuring External Sub-Agents

​Registering the sub agent

​Supporting long-running tasks

​Invoking the External Agent

​Context Management

Overview

The Orchestrator Model

Configuring External Sub-Agents

Registering the sub agent

Supporting long-running tasks

Invoking the External Agent

Context Management