Multi‑modal agents have made notable strides in interacting with open worlds, yet their capacity for self‑evolution remains limited. Supervised fine‑tuning enables imitation but lacks active experience organization, while outcome‑based reinforcement learning encourages exploration but underutilizes intermediate trajectory information. To address this gap we introduce EvolvingBrain, which treats an adaptive skill system as the agent's brain and comprises two principal modules.\ \ Stratified Skill Self‑Evolution extracts procedural patterns and problematic patterns from model‑specific trajectories, synthesizes reusable skills, and iteratively evolves through multiple world interactions. This process yields a stratified skill architecture consisting of domain, plan, and action/trap skills, furnishing the agent with a multi‑granularity knowledge repository and a novel evolution pathway.\ \ State‑Aware Adaptive Execution compresses long‑horizon interactions into a structured abstractive memory and leverages it to drive a state‑aware skill fusion strategy, allowing the agent to flexibly invoke appropriate skills based on its current state.\ \ Experiments on VAB‑Minecraft and DiscoveryWorld demonstrate that EvolvingBrain markedly improves task success rates and exhibits strong generalization across environments. The code is included in the supplementary materials and will be publicly released with the camera‑ready version.\ \ Review: EvolvingBrain systematically extracts and organizes experience, bridging the gap from imitation to autonomous evolution, and offers a practical route for long‑term learning in open‑world agents.