Speaking at API:World
Talk at API:World 2026 on catching unsafe agent behaviour before release and stopping it at runtime, using policy tests, grounded-answer checks, trajectory tests, and versioned tool contracts.
Spoke on the Main Stage Theater at API:World 2026 on 2 September, in a 25 minute best practices session, on building a control layer around an AI agent.
Companies are beginning to let agents search customer data, change settings, and call production tools. A secure model endpoint does not make those actions safe: an agent can choose the wrong tool, use a forbidden parameter, invent a value, or reach a correct answer through an unacceptable sequence of steps. Answer-only evaluation misses all of that, because it grades the destination and ignores the route.
The talk covers four checks that can run in a continuous integration pipeline and at a tool gateway: policy tests for which tools and parameters an agent may use, grounded-answer checks that tie a claim to its evidence, trajectory tests that judge the sequence of steps rather than the final answer, and versioned tool contracts so a tool's interface cannot change underneath a deployed agent. Attendees leave with a pattern they can adapt to their own pipeline and gateway.
