Implement Observation Agent with Claude Sonnet 4 and BentoML Hook

Implement the 'Observation Agent' using Mastra AI. This agent will simulate receiving input from a 'curious AI' camera/drone, which you will model as an inference endpoint deployed on BentoML Cloud (for this prompt, a simple function call will suffice, but conceptualize it as a BentoML service). The agent should use Claude Sonnet 4 to interpret observation data (e.g., 'item detected at location X') and update its internal memory. Integrate a simple tool to 'report_observation' to the Inventory Agent.

```typescript
import { createAgent } from '@mastra-ai/core';

// Simulate a BentoML inference call
async function callBentoMLInference(imageData: string): Promise<{ detected_item: string, location: string }> {
  // In a real scenario, this would be an HTTP call to your BentoML service
  console.log(`Simulating BentoML inference for image data: ${imageData}`);
  return { detected_item: 'SKU12345', location: 'Aisle 3, Shelf 5' };
}

const observationAgent = createAgent({
  name: 'observationAgent',
  llm: 'claude-sonnet-4',
  actions: {
    processObservation: async (ctx, imageData: string) => {
      const inferenceResult = await callBentoMLInference(imageData); // Call simulated BentoML service
      ctx.memory.set('last_observation', inferenceResult); // Update agent's memory
      // Use LLM to interpret and potentially call another tool/agent
      const response = await ctx.llm.chat([{
        role: 'user',
        content: `Interprete this observation: Item ${inferenceResult.detected_item} detected at ${inferenceResult.location}. What should I do?`
      }]);
      // Example tool call or message to another agent based on LLM's response
      // await ctx.send('inventoryAgent', { type: 'item_detected', payload: inferenceResult });
      return response.content;
    }
  },
  // ... other configurations like memory providers, etc.
});

// Example usage: observationAgent.actions.processObservation('drone_feed_001');
```