Resilience: When Tools Fail

Agents depend on the real world. Real-world APIs go down, return 500 errors, or time out. A bad agent will crash. A Great Agent will realize the tool failed and try a different approach or inform the user politely.

1. Types of Agent Failures

• Tool Error: The Lambda function crashed or timed out.
• Hallucinated Tool: The agent tried to call a function that doesn't exist.
• Logic Deadlock: The agent keeps calling the same tool and getting the same error (Looping).

2. Recovery Strategies

• Max Iterations: In Bedrock, you can limit how many steps an agent can take (e.g., 5 steps max). This prevents expensive "infinite loops."
• Error Feedback: If a tool fails, return a clear error message to the agent: "The weather API is currently offline. Please try a different source." This allows the agent's Reasoning to adapt.

---

3. Visualizing the Recovery

graph TD
    T[Thought: Call API] --> A[Action: Tool A]
    A -->|FAIL Error 500| O[Observation: Tool A is offline]
    O --> T2[Thought: Tool A is down. I will try Tool B instead.]
    T2 --> B[Action: Tool B]
    B -->|Success| Final[Goal Met]

---

4. Engineering Tip: Graceful Degradation

If your agent is doing 3 things and the 3rd one fails, should it delete the first 2?

• Recommendation: Always design your agents to return a Partial Success.
• "I found your flight, but the hotel booking system is down. I have saved the flight details for you."

---

Summary

• Failure is a part of autonomous agent life.
• Max Iterations prevent infinite loops and cost spikes.
• Informative Error Messages help the agent's brain pivot to a new plan.
• Partial Success is better than total failure for user satisfaction.