Key Concepts
- 1Agents fail silently and mysteriously. An LLM might hallucinate, a tool might time out, a user might give ambiguous input. Your job as PM is to assume failure will happen and design graceful degradation.
- 2Never show the user a broken experience. If the agent fails, fall back to a simpler system that *works*. Example: if agent times out, show simple search results instead of 'Agent is thinking...'
- 3Tool timeouts: Set reasonable limits (e.g., 10 seconds max for all tool calls). If tools take too long, timeout and fall back to cached/default results.
- 4Hallucination handling: If an LLM invents a product or price, the tool should catch it. Example: agent says 'recommend product ID 99999,' but that product doesn't exist. The tool rejects it, and the agent gets an error. Agent must handle the error gracefully.
- 5Ambiguous input: If a user says 'I want something nice,' the agent should ask for clarification rather than guessing. Example: 'I'm not sure what you mean by 'nice.' Do you mean affordable, premium quality, trending, or popular?'
- 6Monitoring & alerting: Log agent decisions, tool calls, timeouts, and rejections. If rejection rate spikes (users are saying 'no' a lot), alert the team. You have a data quality or model issue.
- 7Graceful degradation: Design the fallback path. Example: (1) Try agentic recommendation, (2) If fails, show top products by rating, (3) If that fails, show search results, (4) If that fails, show 'Browse by category.'
Why It Matters for PMs
Reliability > intelligence: A dumb system that always works beats a smart system that fails. Users prefer a simple 'browse by category' to a broken AI that doesn't respond.
Cost of failure: If an agent makes a bad recommendation and the user returns it, you lose margin (restocking, shipping back, customer satisfaction). A simple fall back cost nothing in comparison.
Trust: Users will forgive a failed agent if you gracefully fall back. They won't forgive an agent that breaks silently (user doesn't realize the recommendation is broken and buys the wrong thing).
Test Your Understanding
Question 1 (Multiple Choice)
What's the best fallback when an agent times out?
Question 2 (Multiple Choice)
If an agent hallucinates a product (recommends something that doesn't exist), how should you handle it?
Question 3 (Multiple Choice)
What does 'graceful degradation' mean in the context of agentic systems?
Question 4 (Short Answer)
Design a graceful degradation path for an agent-powered product recommendation.
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
What metrics would you monitor to detect when an agent is failing?
Think about your answer before revealing the explanation.