Develop Maintenance Predictor with Multimodal Analysis and Adaptive Budgets

Create a 'Maintenance Predictor' agent. This agent should analyze simulated multimodal inputs (logs, thermal images, audio data) using Gemini 2.5 Pro to detect early signs of hardware degradation. Implement an adaptive thinking budget for its analysis: allocate more computational 'thought' for critical or ambiguous signals. Its output should be a predictive maintenance recommendation with estimated urgency.