Design Challenge: A Support Agent That Can Actually Take Actions

Difficulty: Hard · Concept: The 9-Step Framework

Same no-auto-grading, self-assessment format as the RAG design challenge — write a real end-to-end design in free text, then reveal a rubric and the site's own real walkthrough for this exact problem.

The problem: a third shape again. You've been asked to design a system that lets a model do things — look up an order, issue a refund, escalate a case — safely, for 50,000 conversations a day, without either being uselessly restricted or dangerously unrestricted. This is where the hard part stops being about data or infrastructure and starts being about trust.

Design it end to end — how does the agent actually reach real systems, how do you decide which of its actions need a human in the loop versus which can run on their own, and what stops it from being talked into using a capability (like a refund) it never should have used?

Next: Design Challenge: Personalized Recommendations for a 50-Million-Item Catalog — a fourth shape, where the system trains on the consequences of its own past decisions.