- Build intelligent agents with custom tools
- Leverage SAP HANA Cloud for data grounding
- Integrate seamlessly with SAP Cloud SDK for AI
- Inspect agent behavior with Langfuse
At SAP TechEd 2025, a compelling demonstration showcased the streamlined process of building and governing pro-code AI agents, emphasizing flexibility, intelligent grounding, and transparent operations.
The session began by highlighting the freedom developers have in choosing their preferred tools for agent construction. The presenter demonstrated using Cup and Langraph, emphasizing how these tools empower agents with "legs and arms" to interact with various systems. A critical aspect of building intelligent agents is their ability to access and utilize relevant data. This was addressed by integrating an SAP HANA Cloud MCP server, allowing agents to directly ground their predictions based on real-time values from SAP Rapid One, ensuring data-driven accuracy.
Giving the agent its "brain" was simplified through the SAP Cloud SDK for AI, which has been extended for agent development. This SDK provides developers with the flexibility to select from a wide array of models available within the AI Foundation, with a specific mention of a "sonnet 4.5 day," indicating a preference for advanced language models. Once the agent's intelligence was configured, the focus shifted to deployment, a process made efficient for integration into Juul, with the final touch being the A2A agent card.
The demonstration culminated in a live scenario where the newly deployed agent was tasked with optimizing a sales inquiry from a customer named Alinova. The agent swiftly provided promising recommendations and proceeded to create a sales quotation, showcasing its practical application in accelerating business processes. This real-time execution underscored the agent's capability to deliver tangible business value.
To ensure transparency and facilitate debugging, the session introduced Langfuse as a favorite agent development tool. Langfuse allows developers to meticulously inspect all tool calls and the agent's "inner monologue," confirming that optimizations were indeed grounded in the data fetched earlier. This level of insight is crucial for validating agent performance and fostering trust in AI-driven decisions.
“Tools gave the agent legs and arms. But what we are still missing is its brain.”
- Mathis Boerner, Demo Expert
