Our hackathon "Agentic AI Against Aging" has begun Let’s figure out what "agentic AI" in its title means: what (or whom!) are we going to put in the ring against aging? Just in case: what is AI? Simple—in this context, it means neural networks, mainly LLMs. Now to the most interesting part: what is agency? It is an agent’s ability to freely choose a desired goal, the means of achieving it, and, ultimately, to achieve it. The first thing to understand about agency is that it is a spectrum. A spectrum of how much an agent’s goals are truly its own, how free it is in choosing the means and paths to achieve those goals, and how persistent and adaptive it can be. We assume that the AI agents we will build do not have goals of their own, and will do what we tell them. So in this regard they do not have agency. And what about freedom in choosing the means and paths to achieve goals? In this sense, AI already has agency: - We can give an LLM various tools—for example, web search or a RAG knowledge base—and the freedom to choose when and how to use them, as well as whether to use them at all. - At the system prompt level, we can avoid prescribing a specific way to perform tasks and leave the LLM the option to decide on its own how to handle the task. So, strictly speaking, there are easy ways to increase a neural net’s agency, but those who have used them a lot know this doesn’t always help. Often, on the contrary, you need to spell out in maximum detail what exactly the model should do, in what sequence, and how, to achieve the optimal result. That is, sometimes we take agency away from AI to improve the outcome :) But agency truly manifests in multi-agent systems consisting of agents that can decide which next agent to hand off a task or subtask to. Then the system’s behavior evolves from a simple pipeline—where all the steps of solving the task are predefined—into something indeterminate that will “think” not only about the task, but also about the very process of solving it: the type of the next step and how many steps there should be. And with proper tuning, this gives rise to persistence, adaptability, and goal-directedness: the system can try and err, draw conclusions and try again, seek alternative paths, and work as long as necessary until the task is solved. In theory. But in practice, people often call plain LLMs with a system prompt “agents,” even if their capabilities are highly deterministic. And “multi-agent systems” are pipelines or architectures composed of such “agents.” That does not prevent such multi-agent systems from being a very powerful tool if you know what you want and how the step-by-step process of solving your task should look.