Anyone exploring Agentic AI will quickly encounter terms such as «reasoning», «human-in-the-loop» and «Model Context Protocol». This glossary explains the key concepts in just a few minutes.
Generative AI systems such as ChatGPT create text, images, summaries or code. They respond to requests and generate content. Agentic AI, by contrast, pursues a goal, acts autonomously, plans the necessary steps and can trigger actions through connected systems.
For example, an agent can gather information from different systems, prioritise support requests, prepare a quotation or coordinate a customer meeting. Rather than requiring each step to be initiated individually, the system works towards a defined goal and does not necessarily need a request: it can respond to its environment.
Generative AI compared with agentic AI
| Generative AI | Agentic AI |
|---|---|
| answers questions | pursues goals |
| responds to prompts | acts within defined boundaries |
| generates content | executes actions |
| individual interactions | multi-step processes |
Agentic AI therefore supports the completion of tasks. It is based on systems that can retain information and intermediate steps, pursue a goal and access applications, data sources or other tools. This enables them to link multiple work steps rather than merely respond to individual requests. Agentic AI therefore consists of stateful AI systems – the first key term:
Stateful: an AI system works with context
Stateful systems can take relevant information from earlier steps into account. For example, they store task status, decisions, results or other contextual information and incorporate it into subsequent actions. This enables them to handle multi-step processes over an extended period.
Example: An agent remembers which suppliers have already been contacted, which responses have been received and which next steps are still outstanding.
Stateless: an AI system works without stored context
A stateless system treats each request largely in isolation. It does not automatically know which steps have already been completed or which goal is being pursued across multiple work steps.
Example: You ask an AI system to write a summary. For the next request, you need to provide the relevant context again.
Autonomy: how independently can an AI agent act?
While generative AI primarily processes information and generates content, agentic systems can use this information to plan, decide and act autonomously. The term autonomy captures the extent of this independence.
The greater the autonomy, the more tasks can be handled without direct human intervention. At the same time, the need for clear guardrails and oversight increases.
Spectrum of autonomy: different levels of independence
Agentic AI exists in different forms. Some systems support individual tasks, while others can execute entire process steps largely autonomously. Between simple automation and largely autonomous orchestration, there are four levels: reactive automation, task-specific agency, conditional agency and autonomous orchestration.
Learn more about the four levels of autonomy in the white paper «From generative to agentic AI».
In addition to their autonomy, agents are characterised by how they operate. They analyse information, plan the next steps, use external tools and, depending on the use case, collaborate with other agents or people. The following concepts provide a simplified description of how agentic systems work.
PRAF: Perception–Reasoning–Action–Feedback: an agent’s operating cycle
Many agentic systems operate in a recurring cycle:
- Perceive information (Perception)
- Reason and plan (Reasoning)
- Act (Action)
- Review the outcome (Feedback)
This cycle enables agents to respond to new information and adapt their approach.
Communication with external systems and other agents: APIs, MCP and A2A
AI agents do not typically operate in isolation. Through application programming interfaces (APIs) or the Model Context Protocol (MCP), they can access external applications, data sources and tools.
APIs enable targeted connections to individual systems, while MCP provides a standardised framework for accessing them. Put simply, MCP works like a universal adapter: it enables AI agents and enterprise applications to speak the same language and exchange information.
For example, an AI agent can retrieve information from a CRM system, check an appointment in a calendar or create an entry in a ticketing system. This means Agentic AI is not limited to processing information: it can also trigger actions in existing enterprise applications.
Agent-to-Agent communication (A2A) refers to the exchange of information and coordination of tasks between agents. When several specialised agents work together, this is known as a multi-agent system. For example, one agent researches information, another produces the analysis and a third handles the documentation. Anthropic has described further patterns for building agents.
Human-in-the-loop (HITL): people remain in control
With this approach, people remain part of the process. Critical or irreversible decisions are only executed after they have been reviewed and approved by a person. It is important to define exactly when a case must be escalated to a person and which process steps must always remain under human control.
Example: An agent drafts a contract, while a subject-matter expert provides approval.
What Agentic AI can do today – and where its limits lie
Agentic AI is currently surrounded by high expectations. A realistic view of its capabilities is therefore essential. Agentic AI can analyse information, plan tasks, connect different systems and execute defined processes with a degree of autonomy.
At the same time, this is not a technology with consciousness or human judgement. Current systems can make errors and require clear rules, suitable data and – depending on the level of automation – continued human oversight.
For organisations, this means that Agentic AI replaces neither expertise nor responsibility. It delivers the greatest value when people and AI combine their respective strengths. Human oversight, clear guardrails and transparent processes remain key success factors.
Find out more in the white paper
This glossary provides an initial overview of the key terms and concepts. To gain a deeper understanding of how Agentic AI works technically, which security, governance and data privacy aspects need to be considered, and how organisations can plan for production deployment, read the white paper «From generative to agentic AI».
