A paradigm shift is currently underway: ‘generative agents’ use the same models at their core, but extend them with proactive capabilities. They set goals, plan steps, use tools and learn from experience. A study by McKinsey describes the development from pure knowledge tools to agents that automatically perform complex multi-step processes[1].
This strategy paper is aimed at innovation and digitalisation managers, product managers in the technology sector, and investors looking for the next big thing. It offers guidance on the practical introduction of generative agents as personal assistants on edge devices such as AMERIA's MAVERICK AI laptop. The aim is to highlight potential and risks and provide concrete recommendations for implementation.
Definitions: Agents and generative AI
Classic agent concepts
According to Hassan Sawaf of aiXplain, in classic computer science, an agent is a system that perceives its environment via sensors and reacts to it with the help of actuators in order to achieve defined goals. It has internal memory, which allows it to incorporate past experiences into future actions. An agent can act proactively or only respond to stimuli. Agents do not necessarily have to be based on language models; they can be simple rule-based programmes, heuristic methods or even sophisticated systems trained with reinforcement learning.
An ‘agentic AI system’ emphasises goal-oriented autonomy. These systems not only react, but also pursue goals and make decisions independently, using memory structures. Generative AI (‘GenAI’), on the other hand, refers to models such as GPT that generate content such as text, images or audio based on training data. These models work primarily reactively, as they depend on user input. An ‘agentic solution’ combines the autonomy of an agent with the generative capabilities of GenAI. This results in systems that solve complex problems with minimal supervision. Co-pilots, on the other hand, are tools that support humans by making suggestions or automating processes. They remain reactive. An ‘AI assistant’ is also reactive and executes commands without pursuing its own goals.
Differences between chatbots and co-pilots
Chatbots are primarily dialogue-oriented and answer questions. Co-pilots such as Microsoft 365 Copilot provide support with tasks and deliver content, but always respond to user input. Generative agents go one step further: they can initiate, execute and complete tasks independently. They are able to learn from previous interactions and derive decisions. Microsoft describes agents as ‘new apps for an AI-driven world’ and emphasises that they not only provide support, but actually complete tasks[2]. Agents can handle multi-step tasks, such as processing orders, coordinating appointments and analysing data [3]. This requires a combination of generative AI, memory, access rights and tool integration[4].
Advances in generative AI and agentic systems
From language to action
McKinsey describes how technology is evolving from knowledge tools to action systems: agents use large language models to execute complex workflows independently[1]. They can plan tasks, operate tools and cooperate with other agents or humans[5]. These systems manage complex processes that are too variable for rule-based automation[6]. A business process such as organising a business trip – with flights, hotels and rental cars – requires many steps. Agents can coordinate and dynamically adapt these processes[6].
Microsoft explains that agents must have memory, permissions and tools to act autonomously and reliably. Memory is important so that interactions retain their respective context. Permissions regulate secure access to data and applications, and tools enable actions to be performed[4]. These three elements form the basis for proactive action.
Research on generative agents
Research on agentic systems is developing rapidly. In a well-known experiment conducted at Stanford University, ‘generative agents’ were placed in a simulated city. The agents have memory, can observe, plan and reflect. In the simulation, they wake up, prepare breakfast, work, meet up and organise parties. The authors show that such a system can generate credible social behaviour with the help of a large language model and a memory architecture[7].
A later study by Stanford HAI used generative agents to simulate the responses of over a thousand real people to social science questions. The agents were based on interviews and a language model. They reproduced the responses of real people with 85 per cent accuracy[8]. This research shows that agents can not only automate processes, but also realistically replicate human behaviour.
Industrial developments
In addition to large corporations, start-ups are also investing in generative agents. EdgeRunner AI offers a generative platform that provides agents for specific tasks on edge hardware. These agents work without an internet connection to protect sensitive data. The company received $17.5 million in funding in 2025 and is developing hyper-personalised assistants that provide chat, translation, RAG search functions, speech-to-text and code generation independently of an internet connection[9]. By operating locally, the agents are energy-efficient, cost-reducing and increase data protection and reliability in environments with limited connectivity[10].
Use cases for generative agents
Productivity and assistance in everyday work
Agents can automate routine tasks. In the Microsoft world, for example, they take care of creating emails, preparing presentations and conducting research. Specialised agents generate sales leads, check orders, coordinate appointments and relieve employees. Through memory and context processing, they can remember projects over the long term and provide proactive support[12].
This autonomy can be realised locally on devices such as AMERIA's MAVERICKAI laptop. The laptop is an edge device with integrated sensors and AI hardware that offers a three-dimensional user interface. It serves as a personal assistant and recognises gestures to enable touchless interaction. Users talk to an agent that orchestrates various AI services and performs tasks such as scheduling, data analysis and visualisation[13]. Thanks to local operation, data remains on the device and the assistant is available even without an internet connection.
Customer service and knowledge work
In industries such as retail, telecommunications and manufacturing, agents can automatically process customer enquiries and returns and provide support. According to Microsoft, agents can retrieve information from product catalogues and generate personalised responses, thereby improving customer service and increasing productivity.
EdgeRunner AI is aimed at military and industrial users. Its agents offer chat, summaries and translations, but operate completely locally and without a connection. This is particularly relevant for environments with strict security requirements or limited infrastructure[14].
Research and simulation
Generative agents are also used as a tool in research. The Stanford study shows how generative agents can be used to simulate human responses in social experiments[8]. Companies could use such simulations to test products or optimise marketing strategies without revealing real customer data.
Technical requirements
Hardware
Generative agents require powerful hardware. Modern AI laptops such as MAVERICKAI or AI PCs contain special processors with neural engines. Apple describes its M4 chip as having a neural engine with 38 trillion operations per second, which is specially optimised for machine learning[15]. Such processors allow models to be run locally without relying on the cloud. Edge chips such as NVIDIA's Jetson Orin, Hailo-8 or Google's Edge TPU are further examples. They offer between 4 and 400 TOPS (‘tera operations per second’) with low energy consumption[16]. Companies must plan their hardware selection carefully and pay attention to the required performance and energy consumption.
Models and software
Agents are often based on large language models that are optimised for use on edge devices through distillation, quantisation and compression. Compression reduces memory requirements and computing time without significantly compromising accuracy. Knowledge distillation allows a smaller model to learn from the behaviour of a large model. Companies such as EdgeRunner use multiple open-source LLMs that have been adapted for local operation[9].
Memory, security and integration
To work proactively, agents need memory structures that can store past interactions and retrieve them when relevant[12]. At the same time, they must manage permissions to access only the data that is necessary[18]. Security concepts such as trusted execution environments, homomorphic encryption and data anonymisation are necessary to protect sensitive information[19]. Federated learning makes it possible to train models on multiple devices without centralising data.
Implications for industry and the workplace
Productivity gains and new business models
Agentic AI systems promise significant productivity gains. Deloitte predicts that a quarter of companies using generative AI in 2025 will launch agent pilot projects; by 2027, this figure is expected to rise to 50 per cent[20]. Agents can automate complex processes that previously involved a high degree of manual work, for example in software development, purchasing or human resources[21]. This is giving rise to new business models: companies can offer independent agent platforms, sell subscriptions for specific agents or provide services related to training and customising agents. Investors are already pouring billions into start-ups that develop agentic systems[22].
Workplace and qualification requirements
Agents are changing job profiles. Routine tasks are being automated, while humans are increasingly taking on supervisory, creative and strategic roles. Employees must learn to delegate tasks, interpret results and set clear goals for agents. The need for data ethics, data protection and supervision will increase. At the same time, new professional fields will emerge around the development, training and governance of agents.
Opportunities for AMERIA
With MAVERICKAI, AMERIA is positioning itself as a first mover for local generative agents. The device combines powerful hardware with an integrated 3D interface and enables AI applications to run locally directly on the device. For companies, this offers the opportunity to provide sensitive customer service or internal assistance without cloud dependency. Investors can invest in a category that promises high demand due to data protection requirements, energy efficiency and robustness.
Recommendations for decision-makers and investors
- Define pilot projects: Companies should identify specific tasks that are suitable for agentic automation. Coordinating appointments, creating reports or summarising enquiries are good starting points.
- Evaluate hardware: Edge devices with specialised AI processors enable local execution. Decision-makers should check performance, energy consumption and integration[16].
- Use model optimisation: Compression and distillation can be used to adapt models to the capacities of edge hardware[17].
- Prioritise data protection: Data protection-compliant agents require secure storage, access controls and encryption[19]. Compliance with the GDPR is essential.
- Train employees: Managers should offer training so that employees learn how to work with agents, review results and take responsibility.
- Build partnerships: Collaboration with providers such as AMERIA or start-ups such as EdgeRunner enables access to expertise and accelerates the introduction of new services.
- Establish governance: Clear AI governance defines guidelines for use, data security and ethical aspects. An interdisciplinary team should monitor development and deployment.
Conclusion
Generative agents mark the transition from reactive assistants to autonomous digital companions. They use large language models, memory structures and tool integration to independently handle complex processes. Research shows that agents can convincingly simulate social behaviour and are capable of realistically reproducing human actions. Companies that embrace this technology early on can increase their productivity and develop new business models. In combination with edge devices such as the MAVERICKAI laptop, generative agents offer a personal assistant that is available around the clock – even without an internet connection – and adapts to the needs of users[13]. The next few years will show how these systems establish themselves. One thing is clear, however: the transformation has begun, and those who act now will play an active role in shaping the future.
[1] [5] [6] Why AI agents are the next frontier of generative AI | McKinsey
https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/why-agents-are-the-next-frontier-of-generative-ai
[2] [3] [4] [11] [12] [18] AI agents — what they are, and how they'll change the way we work - Source
https://news.microsoft.com/source/features/ai/ai-agents-what-they-are-and-how-theyll-change-the-way-we-work/
[7] [2304.03442] Generative Agents: Interactive Simulacra of Human Behavior
https://arxiv.org/abs/2304.03442
[8] Simulating Human Behavior with AI Agents | Stanford HAI
https://hai.stanford.edu/policy/simulating-human-behavior-with-ai-agents
[9] [10] [14] Newsroom | EdgeRunner Raises $17.5M to Develop Air-Gapped, On-Device AI for the Warfighter
https://www.edgerunnerai.com/news/edgerunner-raises-17-5m-to-develop-air-gapped-on-device-ai-for-the-warfighter
[13] AMERIA AG | MAVERICK AI | Our journey towards a Touchfree future for everyone.
https://www.ameria.com/maverickai
[15] Apple introduces M4 chip - Apple
https://www.apple.com/newsroom/2024/05/apple-introduces-m4-chip/
[16] Top 20 AI Chip Makers: NVIDIA & Its Competitors in 2025
https://research.aimultiple.com/ai-chip-makers/
[17] [19] Optimizing Edge AI: A Comprehensive Survey on Data, Model, and System Strategies
https://arxiv.org/html/2501.03265v1
[20] [21] [22] Autonomous generative AI agents | Deloitte Insights
https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/autonomous-generative-ai-agents-still-under-development.html
-p-500.jpg)
