Traditionally, machine learning algorithms operate in a static environment, like water flowing along a pre-cut channel without changing its shape. The new method flips this paradigm: it lets an agent simultaneously master actions and deliberately modify the very environment it inhabits. Imagine a stream that doesn't just run downhill but actively erodes soft soil, carving an ever shorter and more efficient path. That's how the system shapes its own 'channel' of possibilities.

The key innovation is splitting environment adjustment into two phases, each playing a distinct role in learning. The first phase, akin to the immediate flow, affects the dynamics of the current step: it determines how instant changes in the surroundings influence the agent's actions. The second phase, like slow erosion, shapes the long-term structure of future events: it governs how the environment will respond to actions in subsequent cycles. This separation avoids the chaos inevitable when trying to optimize everything at once and allows precise calculation of how each local change impacts the final outcome. The algorithm needs no ready-made map or predefined rules: it learns purely from its own trial and error, gradually building and improving its world.

The approach was tested on a drone swarm: they simultaneously transmit signals and adjust their positions, much like water finding the shortest path downhill. Notably, trying to optimize all environmental and behavioral parameters at once can degrade the final result — the system must balance immediate gains (quick success under current conditions) against long-term improvements (shaping a more favorable environment). This approach promises to make technology truly adaptive: robots and algorithms could not just react to circumstances but actively mold them to their needs.