A context engineer is debugging a Databricks customer support agent that handles complex billing escalations. The system prompt instructs the agent to: (1) strictly apply the company’s official refund policy; (2) treat entries from the “Billing-Policies” tool as authoritative; and (3) treat FAQ snippets and prior chat history as non-binding guidance only. In one multi-step trace, the engineer observes the following: Step 1: The agent calls the “Billing-Policies” tool and retrieves the official refund policy, which clearly states that chargebacks filed more than 90 days after purchase are ineligible for refunds. Step 2: The agent calls a “Customer-FAQ-Search” tool that returns a community-authored FAQ thread. A top-voted but outdated and unofficial answer incorrectly claims that chargebacks are fully refundable regardless of when they are filed, and the agent summarizes this into the running conversation as an established fact. Step 3: In a later turn, the customer describes a chargeback dispute from 18 months ago. The agent reasons from its own earlier summary, approves a full refund, and shows no reference to the official policy section that excludes cases older than 90 days, even though that policy text is still present in the context. Step 4: In subsequent similar cases, the agent repeatedly approves refunds for old chargebacks, each time citing its prior summary as justification. Which context failure mode describes what is going wrong in this trace?
