From Intent to Evidence: Keeping AI Agents within scope

Duration: NaN:NaN:NaN

Speaker: Siddhant Khare

Type: CLS - Talk (Open Tech Summit)

Expertise:

Event: Devcon

Date: Nov 2026

An agent asked to summarize one ticket may have credentials to read every project and publish anywhere. Its final answer does not reveal which permissions it used. We will follow one task through a permitted read, a blocked out-of-scope request, and a write that requires approval. Then we will correlate execution traces with gateway decisions to distinguish attempted calls, blocked requests, and observed outcomes. The approach combines short-lived, task-scoped permissions with authorization checks before tool execution. Intent interpretation proposes actions. Explicit approval and deterministic checks establish authority. Learn how to constrain agent access and investigate unexpected behavior using open-source examples. We will also cover what incomplete traces cannot establish and why revocation cannot undo an action already executed.