In this article I want to describe a project created by our team within the 48-hour hackathon from JetBrains and Google Cloud in Serbia (06.05.2025-08.05.2025).
Nowadays large IT companies are actively competing in the quality of released AI agents. Now Autonomous Code Generation Agents such as Cursor, Junie, Windsurf Editor and others are becoming especially popular.
Such models are autonomous software assistants based on large language models that can interpret a developer's request, divide it into a sequence of subtasks for code generation, refactoring and testing, and then integrate the obtained changes directly into the development environment. They automatically execute unit tests and CI processes, analyze the build results and make adjustments until all checks are successfully passed.
In our project, we decided to go further and create an entire IT team of such agents.
Why do we expect the agent team approach to open up more development opportunities?
Our system will allow multiple tasks to be performed in parallel by distributing them among specialized agents. For example, while one agent is writing the main code, another agent can do testing and a third agent can do documentation. Large code writing tasks can be split into independent blocks and assigned to several agents in parallel. Such distribution of responsibilities significantly speeds up the development process compared to using a single agent.
Each agent in our team can be customized for a specific role and specialization. We can create agents with deep knowledge in certain development areas, frameworks or programming languages. This allows you to get a better result compared to using a one-size-fits-all agent who has to be competent in all aspects at once.
Using specialized agents allows you to optimize the cost of computing resources. Instead of using one large and expensive model for all tasks, we can select less resource-intensive models for specific specializations. This is especially true for tasks that do not require complex computations, such as formatting code or writing basic documentation.
Working with a team of AI agents opens up a wide range of opportunities to study their interaction and collective decision making. This allows studying aspects such as task allocation, conflict resolution and coordination between agents. Moreover, this approach gives the opportunity to experiment with different team organization models and project management strategies.
We decided that the interaction of our project with the user should take place through a plugin for JetBrains IDE.