-
Notifications
You must be signed in to change notification settings - Fork 65
Description
@awsapm can you check on this issue
Based on my analysis of your system's telemetry data, I've identified the core issue causing your pet clinic agent to recommend nutritional products you don't offer.
Primary Issue: Data Mismatch Between Agent Knowledge and Available Products
Supported Pet Types (200 responses):
• Dog (1,187 requests)
• Cat (1,139 requests)
• Bird (780 requests)
• Hamster (638 requests)
• Snake (619 requests)
• Lizard (404 requests)
Unsupported Pet Types Being Requested (404 responses):
• Rabbit (1,507 requests - highest failure rate!)
• Ferret (369 requests)
• Turtle (367 requests)
• Guinea pig (340 requests)
• Parrot (182 requests)
• Bearded dragon (167 requests)
• Fish (120 requests)
• And many others...
Critical System Issues Identified:
-
Agent Communication Failures:
• Pet clinic agent → nutrition agent fault rate: 67.82%
• 100% error rate on AgentCore.Runtime.Invoke operations
• 25,191 OSError occurrences: "Too many open files" -
Resource Exhaustion:
• File descriptor leaks in the pet clinic agent
• Bedrock throttling exceptions (19 occurrences)
• Extremely high latency (up to 13.5 million ms) -
Agent Hallucination:
• The nutrition agent is generating recommendations for pets not in your database
• Error logs show: "Nutrition service could not find information for pet: turtle"
• But the agent continues to provide recommendations anyway
Immediate Action Items:
-
Fix Agent Configuration:
• Update the pet clinic agent's knowledge base to only include supported pet types
• Configure proper error handling when nutrition data is unavailable
• Add validation to prevent recommendations for unsupported pets -
Resolve Resource Issues:
• Fix file descriptor leaks in the pet clinic agent code
• Implement proper connection pooling for Bedrock clients
• Add circuit breaker patterns for failed nutrition service calls -
Update Nutrition Database:
• Either add nutrition data for the missing pet types (rabbit, ferret, turtle, etc.)
• Or configure the agent to gracefully handle unsupported pets with appropriate messaging -
Implement Fallback Logic:
• When nutrition data is unavailable, provide general pet care advice instead of specific product recommendations
• Add clear messaging about which pets your clinic supports