Put the bowl on the plate, then grab the cup and put it into the bowl that just moved—for a person this is just tidying up the table, but for a robot it is a world rewritten by the previous action. The bowl's position has changed, the arm's pose has changed, and what comes next may be hard to see, hard to grasp, or hard to place accurately. This is exactly the most awkward part of compositional generalization: individual skills are already quite reliable, but chaining them together collapses.
Recursia's entry point is not to train yet another stronger policy, but to ask how improvement itself is organized. It splits the system into an inner and an outer loop: the inner Agent handles observation, decision-making, and code execution within the current task; the outer Agent aggregates execution experience, proposes, validates, and retains modifications to the shared stack, letting useful changes enter subsequent loops.
Here we first need to clarify a question of definition. The article adopts a stricter definition of RSI: improvement must not only be deposited as reusable memory, skills, or tools, but should also enhance the system's ability to discover, design, and validate the next round of improvement. In other words, true recursive self-improvement asks whether "AI can improve the mechanism that improves AI," not merely whether "this task was done better." This distinction determines how all the numbers that follow should be read.