Systems offlineAdvanced
The AI agent gone wrong
Your customer-support agent handles refunds on its own. This week it looped for half an hour, ignored what users told it, emailed a customer three times, tried to delete records and refunded money it shouldn't have. Seven parts of the agent each have a problem. Pick any, read its trace, and choose the fix.
Online 0 / 7Down 0Not checked yet 7XP earned 0
Pick a component
Each component has one question, so read it carefully. Right answers bring it back online.
What users can do
- DownFinishes within its step and cost budgetNeeds Agent loop, Model calls
- DownRemembers what the user saidNeeds Conversation
- DownTakes actions safelyNeeds Action tools, Page reader
- DownHands valid data to billingNeeds Result handoff
- DownCan be debugged when it's wrongNeeds Run records
What this incident covers
- Agent loop (plans and acts): Agent Reliability & Evals
- Conversation (what the model sees): Agent Memory Systems
- Action tools (email, refunds): From Text Generation to Action
- Page reader (fetches web pages): Defending Against Prompt Injection
- Result handoff (to the billing system): Structured Output from LLMs
- Model calls (every step): Understanding AI App Costs
- Run records (what happened): LLM Observability & Tracing