This research aims to enhance the combinatorial generation of assemblages made of spatial units with adaptive control over climate-related factors (such as daylight hours access on exposed surfaces) and topological aspects using Reinforcement Learning (RL). The approach uses RL-trained agents to adapt to lighting conditions in the process of selection and aggregation of parts. The agents move within a voxel space, and place parts (single voxels) aiming to ensure topological consistency and a target daylight hours access on exposed surfaces in the final assemblage. Agents’ behavior is shaped by both intermediate and final rewards, providing them with feedback on the quality of their individual and overall choices. The research introduces condition-based adaptivity into a combinatorial process by means of RL training, moving beyond both simple random choice and predetermined heuristics sets; although both policies can relate to boundary conditions, they are respectively non-controllable and tied to a specific environmental scenario. Through the agent’s trained policy, the system learns a state-action-reward relationship in a continuous feedback process between space, environment and climate data that generalizes to any environmental configuration that can be coded in the system’s terms.