SAP has published a detailed guide for building AI agents using its Cloud Application Programming (CAP) model, demonstrating how enterprises can integrate intelligent systems directly into their existing business applications rather than bolting them on as separate tools. The approach combines language models with domain-specific tools and data, allowing agents to iteratively solve complex problems while staying connected to company databases and services.

What Exactly Is an AI Agent, and Why Should Enterprises Care?

An AI agent is fundamentally an AI model placed in a loop with a set of tools and a clear objective. Unlike a simple chatbot that answers questions once and stops, an agent can reason about what it needs to do, decide which tools to use, and then iterate based on the results it gets back. Think of it like a travel assistant that doesn't just tell you about trips, but can actually search your company's trip database, check availability, and refine its recommendations based on what it finds.

The key difference in SAP's approach is that these agents are built directly into enterprise applications using CAP, rather than existing as separate systems. This means an agent can access your company's real data, understand your business rules, and integrate seamlessly with existing workflows. The agent controls the interaction between the language model, the tools, and the user, feeding tool outputs back into the model as context for the next step until it reaches a final answer or hits a stopping condition like a maximum number of iterations.

How to Build an AI Agent Using SAP's Cloud Application Programming Model

• Set Up Your Environment: Initialize a new CAP project using the cds init command with Node.js runtime, then connect to SAP AI Core service using the sap-ai-sdk package, which provides access to language models like Mistral Small and Claude.
• Define Your Tools: Create tools that perform specific actions, each with a function that does the work and a schema that tells the agent how to use it, including a name, description, and input parameters validated with Zod.
• Create the Agent Loop: Use LangChain's createAgent function to combine your language model with your tools and a system prompt, then invoke it with messages that include both system context and user input.
• Integrate with Your Data: Connect your tools to actual CAP data models and services so the agent can query databases, retrieve real information, and perform actions like searching trips by date range or processing transactions.
• Test and Iterate: Run your agent with test scripts using cds bind to ensure proper environment variables are set, then refine based on how well the agent handles multi-step reasoning tasks.

Why Building Agents Inside Enterprise Systems Changes the Game

The traditional approach to AI agents treats them as external services that companies bolt onto their existing systems. SAP's framework flips this by making agents native to the application layer. This matters because agents need context to make good decisions. When an agent is built directly into a CAP application, it has immediate access to your company's data models, business logic, and existing services without needing to call external APIs or deal with integration complexity.

Consider a practical example: a travel booking agent built with CAP can search your company's trip database by date range, understand your business rules about which trips are available, and provide recommendations based on actual inventory. The agent doesn't need to be told what data exists or how to access it; it's already wired into the system. This reduces latency, improves accuracy, and makes the agent more trustworthy because it's working with authoritative data sources.

The framework also addresses a critical gap in how many companies approach AI. Rather than treating agents as replacements for existing systems, CAP allows them to augment and enhance what's already there. An agent can handle the reasoning and decision-making while the underlying CAP application handles data persistence, security, and compliance. This separation of concerns makes it easier for enterprises to adopt agentic systems without rearchitecting their entire technology stack.

What Tools and Models Are Available for This Approach?

SAP's integration with AI Core provides access to multiple language models through a single interface. Developers can choose from models like Mistral Small for cost-effective testing, Claude for more sophisticated reasoning, and other options depending on their use case and budget constraints. The sap-ai-sdk package handles the connection, making it straightforward to swap models without changing application code.

The framework also leverages LangChain, an open-source library that has become the de facto standard for building agent applications. LangChain provides the createAgent function, message handling, and tool integration patterns that make it easier to build agents without starting from scratch. By combining SAP's enterprise integration capabilities with LangChain's agent patterns, developers get the best of both worlds: enterprise-grade data access and proven agent architecture.

The Bigger Picture: Why This Matters for Enterprise AI

The release of this guide signals a shift in how enterprises should think about AI agents. Rather than viewing them as experimental tools or marketing gimmicks, companies can now see them as a way to make their existing systems smarter and more responsive. An agent built into your ERP system can handle complex queries that would normally require human intervention. It can search multiple data sources, understand context, and provide recommendations based on your company's specific rules and data.

This approach also addresses a common pain point in AI adoption: the gap between proof-of-concept and production. Many AI pilots fail because they don't integrate with real business processes and data. By building agents directly into CAP applications, enterprises can move from experimentation to production more quickly, because the agent is already connected to the systems that matter.

For organizations already using SAP systems, this framework offers a clear path forward. Rather than waiting for new AI-native platforms or struggling with integration challenges, they can start building agents today using tools and patterns they already understand. The tutorial provided by SAP includes working code examples, step-by-step instructions, and practical guidance on connecting to AI Core, defining tools, and testing agents in a real application context.