What Your Agent Does When It Cannot Decide

Most agent demos stop at the happy path. Pampa starts where the happy path ends.

Pampa Dutta, Product Leader at Amazon, begins with one real exception, the kind that does not fit the workflow. Live, she gives an agent its tools and draws the line around what it is allowed to decide on its own, runs it against a case it can close, then runs it against one it cannot, and shows exactly what reaches a human and in what form. She builds in Kiro and maps the setup back to Claude Code as she goes, so you can rebuild it in the stack you already use. One boundary, two runs, screen shared the entire time

What you’ll walk away with:

1. Where the agent’s authority stops is the real architecture. The boundary around what it may decide alone determines which tools it needs, what evidence it keeps, and the moment it halts. Get that line wrong and nothing downstream holds.
2. A handover is a deliverable, not a fallback. What the agent passes to a person has a format: the case, the actions already taken, the evidence gathered, and the single decision it is asking for. Build it as deliberately as the resolution path.
3. The cases that break your agent are the ones you never modelled. Happy-path demos pass because the inputs were picked to pass. Running the same agent against a case it cannot close is the only test that tells you whether an exception path exists at all.