What’s the Concept Behind Agents?
AI agents have taken the tech world by storm. Over the past year, systems that can plan tasks, use tools, and autonomously execute workflows have evolved from research demos into everyday products. Some of the most visible examples can be found in software development. In this blog post series, Niklas Frühauf, Senior Data Scientist at sovanta, explores whether agents are already transforming business processes – or if they are still largely a promise waiting to materialize. To kick things off, he explains the fundamental concepts behind AI agents.
The rise of “vibe coding” describes a workflow where developers describe what they want and AI generates and iterates on the code. Tools like Cursor IDE, Claude Code from Anthropic, or OpenCode illustrate how quickly agentic systems are changing programming workflows. A second wave of adoption has emerged around personal assistants. AI systems (such as ChatGPT, Claude, OpenClaw and others) increasingly act as task managers, research assistants, and workflow helpers capable of coordinating multiple tools and services.
The excitement has fueled bold predictions about the future of work. Studies from organizations like McKinsey & Company and the World Economic Forum estimate that between 15–35% of work activities could eventually be automated by AI technologies, while hundreds of millions of jobs may be reshaped over the coming decade. (Source: www.tiaa.org / www.mckinsey.com)
Yet outside coding environments and consumer assistants, the picture is less clear. While agentic AI dominates headlines and product launches, enterprise adoption beyond pilots and prototypes appears to lag behind the hype, raising an important question: are AI agents already transforming business processes, or are they still largely a promise waiting to materialize? And more importantly: how can SAP customers already profit from AI agents right now? Let’s start with the concept behind agents.
What are AI Agents, exactly?
Before answering that question, it helps to take a step back and clarify what we actually mean when talking about “AI agents”. The term is often used loosely to describe very different types of systems that operate at different levels of autonomy.
A useful way to understand the evolution is to distinguish between three levels of AI-powered applications: simple generative AI, tool-using systems, and autonomous agents.
- Generative AI applications
The simplest form uses a large language model purely for content generation: you give it a prompt and it produces text, code, or images. Imagine asking an AI to invent a new pasta recipe. The model might generate something creative – but it could also produce something unrealistic or terrible, because it is essentially improvising based on patterns in its training data. This makes it powerful for brainstorming or drafting content, but it has no grounding in external information. - Tool-using AI (“chains” or agentic workflows)
The next step adds tool or function calling, allowing the AI to access external systems such as databases, APIs, or search engines. In the cooking example, instead of inventing a recipe from scratch, the AI could search a recipe site such as BBC’s food database, retrieve the top three pasta recipes, and summarize them. These workflows are often called AI chains or agentic pipelines. However, the sequence of steps is still defined by a developer rather than decided autonomously by the model. - Autonomous AI agents
The most advanced level introduces planning and autonomy. Instead of executing a fixed workflow, the system can decide which actions to take, use multiple tools, and iterate until a goal is achieved.Returning to the cooking analogy, an AI agent could check what ingredients are in your fridge, ask about dietary preferences, search for recipes, order missing ingredients, and guide the cooking process. The key difference is that the system plans and adapts its actions dynamically rather than following a predefined chain.
In practice, most “AI agents” used in enterprise environments today sit somewhere between the second and third category: structured workflows with limited autonomy, rather than fully independent digital coworkers.
Use the simplest approach that solves the business problem
An important observation is that as the autonomy and capabilities of these approaches increase, so do their complexity, token consumption, development effort, and audit requirements.
The key takeaway: use the simplest approach that solves the business problem. Not every use case requires a fully autonomous agent.
- For simple tasks such as summarization, classification, or information extraction, a plain LLM call is often sufficient.
- For structured processes that include language understanding, LLM-powered workflows or chains are usually the better fit.
- For dynamic environments with many possible tasks requiring reasoning and planning, AI agents can add real value.
And importantly, generative AI is not the only tool available. Classical machine learning and traditional process automation remain highly effective solutions for many business problems, and in many cases they are still the more reliable and cost-efficient choice.
Different Types of AI Agents
AI agents are heavily dependent on the tools they can access and the knowledge they have about the business processes and platforms they operate within. This has led solution (and platform) providers to take three main approaches:
- Pre-built, platform-specific agents – Many vendors now offer hand-crafted AI agents designed to work out of the box for key business functions (or expose functions via MCP/A2A for easy agent consumption).
- Low-code configurability – Increasingly, business experts can configure custom AI agents using low-code platforms such as n8n or Joule Studio, enabling tailored workflows without deep programming skills.
- Custom pro-code agents – Some tasks still require full developer involvement, particularly when complex reasoning, task-specific tools, or integration with specialized systems is needed.
Understanding these layers is crucial: not every AI agent needs to be built from scratch, but the right approach depends on the task complexity, available tools, and level of autonomy required.
The right approach?
In the next part of this blog post series we focus on the SAP approach to agents.
