Bring Your Own Agent: Integrating Our Chatbot Summy as a Pro-Code Agent in SAP Joule

Three prerequisites had to be met for the standardized use of Summy in SAP Joule: integration with SAP Joule as the central user interface, standardized communication between AI agents, and the use of SAP BTP’s modeling and security services. Without standardized agent communication, the architecture could not be scaled.

Technically, several components had to work together to achieve this: the A2A SDK, authentication via SAP BTP, and declarative configuration on the Joule side.

In addition, there was a key requirement for the architecture: It should cover the current use case but also support future multi-agent scenarios so that additional agents could be integrated later without requiring a redesign.

Recently, it has become possible to integrate custom Pro-Code agents – developed, for example, in Python – into SAP Joule. We chose this approach for Summy: The agent remains a standalone service with its own business logic, but SAP Joule serves as the central user interface. Integration into SAP Joule is handled via A2A, allowing Summy to make its capabilities available directly within Joule.

Within the agent, LangGraph orchestrates the processing steps of a query, from interpreting the question to querying the knowledge base and formulating the response. The Large Language Models (LLM) are connected via SAP AI Core. This enables centralized model management, transparent governance, and secure authentication within SAP BTP. The SAP HANA Cloud Knowledge Graph Engine serves as the knowledge base, in which employees, skills, projects, and references are represented as a knowledge graph. For more information, see the Summy Chatbot Success Story.

At the heart of the integration is a dedicated A2A server based on the A2A protocol. This server publishes an Agent Card – a machine-readable profile containing the agent’s name, capabilities, supported input and output formats, and security requirements. SAP Joule reads this Agent Card and thereby knows what the agent can do and how to communicate with it.

Every call from Joule includes an authentication token from the SAP BTP Cloud Identity Service. The agent server checks the token against the BTP Identity Server and validates its validity. Only the Agent Card is publicly readable; all other endpoints are protected.

Responses are not generated in a single step. The agent reports each completed processing step as an interim result, so that users can see at any time in Joule what is currently being worked on. The agent also returns the finished response as a separate A2A artifact—a clearly addressable result object that Joule specifically retrieves and displays as formatted text.

In SAP Joule Studio, the capabilities of the Joule agent are described using the two concepts “scenario” and “function.” The scenario works similarly to the Agent Card and defines key parameters of the Joule agent, such as its name and the associated function scenario. The function then describes the agent’s executable logic, including sending messages and calling APIs.

Our chatbot, Summy, is fully integrated into SAP Joule! Users can ask questions about projects, employees, and skills directly in SAP Joule, without needing a separate user interface or tool.

This reduces the effort required for searching: A single query provides information that previously had to be compiled across multiple systems. The intermediate results during processing also make it transparent and traceable how an answer is generated and which steps the agent is performing.

The A2A standard provides the foundation for a scalable multi-agent architecture. Additional agents—such as those for bid analysis, consultant matching, or resume creation—can be integrated without modifying the existing architecture. This makes the architecture reusable for many other AI use cases, and in addition to low-code Joule agents, fully custom-developed AI agents can also be seamlessly integrated into Joule.