How an AI agent works: A look under the bonnet of reasoning and multi-agent systems

While conventional AI solutions often seem like a “black box” that produces unpredictable answers, modern AI agents work completely differently. They follow structured thought processes, systematically collect information and make well-considered decisions. At octonomy, we have developed a multi-agent system that implements precisely these principles – and achieves an impressive response quality of 96%.

The difference between standard AI and intelligent reasoning

Standard AI systems often work on the principle of “one-shot answering”: they receive a query and immediately generate an answer based on probabilities. The problem? These answers can be inaccurate, incomplete or even hallucinated.

Our approach at octonomy is fundamentally different. Our AI agents go through a structured reasoning process that starts with a thorough context analysis. The agent first analyses the query and identifies the specific context before systematically navigating through relevant areas of knowledge. Instead of immediately producing an answer, the agent asks specific questions as required to gather all relevant information. Only after this complete collection of information is an informed response produced.

The architecture of our multi-agent system

Central knowledge base as the foundation

At the heart of our system is an intelligently structured knowledge base that is fed from various source systems. Documentation systems such as Confluence and SharePoint provide structured information, while manuals and technical specifications provide detailed knowledge. Workflows and process documentation provide contextual links, and transactional systems enable access to real-time information.

However, what makes our solution special is not just the data collection, but the intelligent processing. The data is not simply stored, but transformed using specialised algorithms. Intelligent chunking breaks down large documents into meaningful units, while normalisation and hierarchical provision create a context-sensitive knowledge graph. This structure enables agents to generate precise and relevant answers instead of getting lost in a sea of unstructured information.

Specialised agents for different tasks

Instead of a generic “jack of all trades”, we rely on specialised agents that are optimised for specific tasks. Support agents focus on processing customer tickets quickly and accurately, and are trained to translate complex technical problems into comprehensible solution steps.

Consultancy agents, on the other hand, are designed for comprehensive field service enquiries. They can access extensive product information and develop customised solutions that are tailored precisely to the needs of the respective customer. These agents not only understand the technical specifications, but can also explain complex relationships between different product components.

Supervisor agents act as intelligent switches in the system. They analyse incoming requests, evaluate their complexity and context and forward them to the best specialist in each case. This orchestration ensures that each enquiry is processed by the agent best suited to the specific problem.

Intelligent tool integration

Our agents are more than just chatbots – they are active system participants that can interact with the entire IT landscape. Integration with ticket systems such as Jira, ServiceNow or FreshDesk enables agents not only to answer enquiries, but also to create, update and track tickets.

Logistics systems provide real-time information on deliveries and order status, while ERP systems provide detailed customer information and order histories. One particularly innovative enhancement is the telephony integration, which makes it possible to hold natural language conversations. Customers can simply call the AI and discuss their concerns as if they were speaking to a human employee.

How reasoning works in practice

Let’s take a concrete example from our Devolo Wi-Fi support to illustrate how our reasoning approach differs from conventional AI systems.

If a customer writes “My Wi-Fi is not working properly”, a standard AI solution would typically output a generic list of troubleshooting steps. This answer may be technically correct, but often does not help to solve the specific problem.

Instead, our octonomy agent starts with a systematic contextualisation. Which specific Devolo product is the customer using? What specific symptoms are occurring? When did the problem first occur? In parallel, the agent searches the top 50 most relevant document areas in our knowledge base, utilising advanced retrieval mechanisms.

The ability to analyse images in real time is particularly impressive. If the knowledge base contains technical diagrams and schematics, the agent can interpret them and include them in its response. An intelligent reranking system then filters out the 10 most accurate sources of information based on the specific context of the enquiry.

The structured response is then a customised step-by-step guide tailored to the customer’s product configuration. If the first answer is not sufficient, the agent enters into dialogue: “Have you already carried out step X? What result did you get?” – and adapts its recommendations accordingly.

how an ai agent works

System prompting: the DNA of the agents

System prompting gives each agent a clear “personality” and way of working. These prompts not only define the agent’s behaviour, but also their limits and quality standards. A typical system prompt could read: “You act as a support AI for area X. You should strictly adhere to the manuals, not make anything up and not use forum knowledge. Be precise and follow up on any ambiguities. For complex cases, gather all relevant information before giving an answer.”

These prompts are highly specific and are carefully customised depending on the use case. An agent for quick ticket processing behaves fundamentally differently to an agent for complex consultations, even though both are based on the same technology.

The future: autonomous full automation

Our system follows a clear development strategy, which we refer to as a maturity curve. The first stage is about knowledge consolidation – the centralised provision of all relevant information in a structured form. Teams can already work much more efficiently in this way, as they no longer have to laboriously navigate through different systems.

The second stage involves assisted automation. The system generates pre-answers and assists with processing, while human experts retain final control. This phase enables teams to significantly increase their productivity without increasing the risk of errors.

The target scenario is autonomous full automation, in which the system processes standard enquiries completely independently. Complex cases are recognised intelligently and forwarded to human experts, who can then concentrate on the really challenging tasks.

Why 96% response quality is possible

The high precision of our answers is the result of a combination of several mutually reinforcing factors. Structured reasoning prevents impulsive answers and ensures well thought-out solutions. The context-sensitive knowledge base with hierarchical information structure enables the agents to find the relevant information even for complex queries.

Specialised agents for specific use cases ensure that each agent is optimally matched to its task. Continuous validation through systematic enquiries ensures that the solution really fits the problem. Integration into existing systems provides real-time information, which is essential for precise answers.

Conclusion: intelligence through structure

Modern AI agents do not “think” like humans, but they follow structured processes that lead to more reliable results than conventional AI systems. At octonomy, we have translated these principles into a practical multi-agent system that not only automates complex support processes, but also delivers exceptional quality.

The result: teams can breathe easy while productivity increases – just as our claim promises: “When your team grows without getting bigger.

Want to learn more about our multi-agent technology? Contact us for a personalised demo and experience how structured AI reasoning can revolutionise your support processes.

Frequently asked questions about ai agents

he “Big 4 AI agents” are often referred to as the leading conversational AI systems that currently dominate the market. These include OpenAI’s ChatGPT, which has become known for its versatility, Anthropic’s Claude, which is valued for its thoughtful and safety-oriented responses, and Google’s Gemini and Microsoft’s Copilot, which are each integrated into their extensive ecosystems.

These four systems have established themselves as particularly powerful in different areas – from text generation to programming and data analysis. Each of these agents has developed specific strengths and is continually evolving to handle more complex tasks.

However, it is important to understand that this market is developing very dynamically. New providers and specialized solutions emerge regularly, and the definition of the “Big 4” can change accordingly. It is therefore crucial for companies not to rely solely on well-known names, but to evaluate the specific requirements and the quality of the results.

Intelligent agents are classified in computer science according to their mode of operation and their capabilities. Simple reflex agents react directly to current perceptions according to pre-programmed rules. They make decisions based only on the current situation, without regard to history or future consequences.

Model-based reflex agents extend this concept by building and maintaining an internal model of the world. They can act meaningfully even if they cannot perceive all relevant information directly, as they derive missing details from their model of the world.

Goal-based agents pursue specific goals and can evaluate various alternative courses of action. They plan their actions in such a way that they optimally achieve their goals, even if this requires several steps. These agents can solve more complex problems where the direct path to the goal is not obvious.

Utility-based agents go one step further and can weigh up different goals. They use a utility function to evaluate how desirable different states are and make decisions that maximize the expected utility.

Finally, learning agents can improve their performance over time. They analyze their experiences, learn from successes and mistakes and adapt their strategies accordingly. This ability to self-improve makes them particularly valuable for complex, changing organizations.

An AI agent is an autonomous software system that can make decisions and carry out actions independently in order to achieve certain goals. In contrast to conventional software, which only responds to direct commands, AI agents can independently perceive and analyze their environment and act accordingly.

a, ChatGPT can certainly be considered an AI agent, but it is a specific type with certain limitations.

ChatGPT fulfills the basic criteria of an AI agent: it perceives its environment by processing and understanding user input, makes autonomous decisions about how to respond, and executes actions by generating relevant and coherent responses. In doing so, it pursues clear goals such as helpfulness, accuracy and engaging communication. This autonomy is demonstrated by the fact that ChatGPT itself decides how to structure and formulate responses.

However, ChatGPT is a relatively limited AI agent. It only responds to requests rather than acting proactively, has no persistent memory between different conversations and cannot perform external actions such as searching the internet or controlling other systems. Every conversation starts anew for ChatGPT.

ChatGPT can best be classified as a conversational AI agent or dialogue agent that specializes in natural language interaction in a limited environment – the chat interface. While it is less comprehensive than autonomous vehicles or trading bots that continuously monitor and respond to their environment, it is significantly more advanced than simple rule-based chatbots that only play back pre-programmed responses.

ChatGPT thus demonstrates the core characteristics of AI agents – perception, decision-making and goal-oriented behavior – but operates within a narrower scope than agents that can operate in the broader digital or physical world.