Multi-agent AI for simulation-based learning of soft skills

The Upbeat project strives to upskill young immigrant entrepreneurs with AI tools that help them build resilience when starting a business in Finland and Estonia. In addition to taking part in a structured training programme, participants were able to test their soft skills capabilities.

Haaga-Helia’s role in the project is to train the trainers and think of new pedagogical methods they can use to upskill the learners. One of these methods was a multi-agent soft-skills simulation game, which was provided by Upbeat’s Estonian trainer Ilya Vasilyev, who is also the co-founder of a start-up called Cyborgs. Because of this partnership, the participating entrepreneurs were able to try the simulation during one of their training sessions in Tallin and in Helsinki.

From chatbots to characters

When thinking about artificial intelligence in corporate training, it usually revolves around coding tutors or adaptive learning platforms. But not all challenges in the workplace are about technical ability. Many are about people: managing conflict, building empathy, navigating cultural differences, and leading under pressure.

In addition, most educational technologies and pedagogical methods approach AI as an assistant: a tool that provides hints, explains concepts, or grades answers. Simulations flip this logic, as AI is not a helper but a character in a workplace drama.

For example, in Cyborgs’ simulation tool, built on a range of GPT models, the characters have long, detailed prompts that define their personalities, motivations, and even hidden agendas. They are explicitly instructed with ‘You are not a helpful assistant. You are in character’.

This distinction matters to the pedagogical process. In simulations, the goal is not to give correct answers but to generate realistic, sometimes challenging interactions that mirror real-world difficulties. A colleague might misinterpret an email. A manager might give vague or unhelpful feedback. A team member might lose their temper. Learners, such as the entrepreneurs in the Upbeat project, experience these scenarios and practise responding – sometimes failing, sometimes succeeding.

Multi-agent systems at work

Technically, the Cyborgs platform is a multi-agent system (MAS). In my interview with Ilya Vasilyev in september 2025, he explained that each character is an autonomous agent, interacting not only with the learner but also with other characters. Conversations evolve dynamically as agents exchange context, share private knowledge, and pursue goals that may align or conflict. This ensures that workplace scenarios do not feel scripted. Instead, they unfold unpredictably, much like in real life.

This architecture reflects broader advances in MAS research, which emphasise collective intelligence emerging from interactions among multiple AI agents. Multi-agent systems allow individual agents with their own properties to collaborate (or compete) in ways that create richer, more realistic outcomes than any single agent could achieve (Sun et al. 2025).

Vibe coding and context engineering

Building such a system requires a distinctive workflow. Companies like Cyborgs embrace vibe coding: delegating entire coding tasks to AI without inspecting every line. The process is iterative, if the output fails, the AI is instructed to fix it until it works. This approach dramatically accelerates prototyping (Chowdhury & Mann 2025).

In the past, start-ups relied on design mock-ups in tools like Figma. With vibe coding, companies can create working prototypes in hours, allowing real users to interact with them almost immediately. This fast feedback loop is invaluable for testing new ideas.

But vibe coding is not the whole story. Vasilyev tells me, that for production-level features, Cyborgs shifts to AI-enhanced programming, where humans read, refine, and own the generated code. Here, the critical skill is not writing syntax but context engineering: providing the AI with precise requirements, constraints, and examples so that the output is robust and maintainable. He adds, that the combination of fast, exploratory vibe coding and careful, structured context engineering gives the four-person team extra agility. They can move quickly without losing control of their core systems.

Cutting-edge AI benefitting pedagogical training

One of the choices any vibe coding startup must make is the strategic choice to avoid dependence of any single AI provider. When the company’s value proposition is built around scenario design and pedagogy, they can benefit from every new advance in the AI ecosystem. It also ensures that improvements in language models, audio synthesis, or even real-time video generation can be integrated quickly to make simulations richer.

From a technical perspective, Cyborgs is a case study in how cutting-edge AI can be applied to human challenges and used as a novel pedagogical tool in training. It combines the unpredictability of multi-agent systems, the speed of vibe coding, and the discipline of context engineering to create something far more engaging than another chatbot.

Ultimately, the aim in the Upbeat project is not to build smarter machines, but rather to help our young entrepreneurs to practise being smarter, more resilient, and more empathetic in the harsh business world.

Picture: Shutterstock