A collaborative effort among teams from Beihang University, The Chinese University of Hong Kong, and the National University of Singapore has yielded the OmniHarness system. This innovative system empowers visual agents to engage in autonomous practice, inspect results, and save effective processes, thereby formulating reusable strategies. In the realm of ComfyBench creative tasks, OmniHarness has demonstrated an impressive 95% resolution rate, outperforming the top control group by a significant margin of 27.5 percentage points. When paired with Codex GPT-4o, the system's overall resolution rate climbs to an equally remarkable 92.5%. Through the application of symbolic strategy learning, OmniHarness systematically organizes successful processes into comprehensive strategies that encompass workflow templates, applicable conditions, and more. Upon subsequent invocation, the system first assesses adaptability before binding new inputs. During its autonomous practice phase, OmniHarness selects topics based on their novelty and the agent's capability boundaries. It conducts multi-round checks and repairs during execution, leveraging both successes and failures as a foundation for future actions. After adapting the strategy library to other agents, OmniHarness can elevate the task resolution rates of multiple systems. Moreover, its capabilities extend to video tasks, achieving an overall video resolution rate of 85.9%. However, elements such as autonomous practice do increase computational costs, indicating that there is still potential for optimization.