SAP has recently released a detailed reference architecture for AI Agents at SAP Architecture Center.

These reference architectures are great. However, for a beginner this could be a bit overwhelming to understand.

In this blog series, we will try to simplify the SAP AI Agent architecture, how the entire call flow works, how Joule Agents interact with 3rd-party agents, how the custom agent works etc.

The #1 blog focuses on…

High-Level AI Agent Architecture

                                                                   Ref: SAP Architecture Center

1. Joule at the Center

Joule is the central orchestrator and entry point for all user interactions.

Joule UI
Joule UI is where users interacts with Joule and enter the prompt.

Joule Orchestrator
Joule’s planning and reasoning engine determines the best way to fulfill it.
Joule orchestrator decides whether the request should be handled by:
• A Joule Skill
• A Joule Agent
• A knowledge retrieval process
• A custom agent

Joule Capabilities
Joule’s common capabilities are Informational, Transactional, Navigational and Analytical.
Joule orchestrator decides which capabilities to be used or delegate the task to an agent.

Joule Skill and Joule Agents
A Joule Skill is a reusable business function that performs a specific task when invoked by Joule. For example, Create Purchase Order or Approve Leave Request.

A Joule Agent is an AI agent. It can plan, reason, and execute multiple actions to achieve a business goal.

For example user may ask “Onboard a new employee joining next Monday”.

In this case, the agent may

• Create employee record
• Assign laptop
• Request access permissions
• Schedule orientation
• Notify manager etc.

A user’s request may involve using a built-in skill or retrieving information or delegating the task to an AI agent.

2. Custom Agents on SAP BTP

SAP supports two complementary paths for building AI agents.

1. Joule Studio based Agents
Build agents using Joule Studio using intent-based development

For Joule Studio Based Agents, the platform handles the integration work. When a user’s prompt matches the agent’s purpose, Joule invokes the corresponding agent.

3. MCP Gateway in Integration Suite

MCP Gateway is going to be a new capability within SAP Integration Suite.

Customers may use MCP Gateway to expose SAP and non-SAP APIs as an MCP server.

It would also help customers, create, manage and expose their own MCP servers along with lifecycle management, governance and monitoring.

For example, a customer needs to expose Purchase Order API from S/4HANA. They may use MCP Gateway to expose these APIs as MCP tools for AI agents.

Check the roadmap to know about its availability.

4. Agent Gateway

Agent Gateway that enables external clients and applications to consume Joule Agents through standardized protocols.

The Agent Gateway exposes Joule Agents via the A2A protocol with an externally reachable endpoint. This enables third-party AI Agents, applications, partner systems and custom business applications to consume SAP-native agents.

For example, using Agent Gateway, 3rd party platform like Microsoft Copilot Studio can delegate SAP-specific tasks to Joule Agents.

The Agent Gateway is not yet generally available.

2 Important Concepts to Revise

1. Agent2Agent (A2A) Protocol
The Agent2Agent (A2A) protocol is an open standard for communication and collaboration between autonomous AI agents.

A2A protocol enables one agent to delegate tasks to another, inquire about its capabilities and exchange information in a structured manner.

2. Model Context Protocol (MCP)
MCP is an open standard that defines how AI models and agents can discover, understand and interact with external tools.

MCP acts as a universal adapter, allowing agents to consume tools without needing to know their underlying implementation details.

What’s Next?

In the next blog, we will simplify the end-to-end call flow… and see what really happens after a user submits a prompt.

Disclaimer: This blog reflects my personal understanding, interpretation, and point of view.