Uploaded July 2026 | Updated September 2026, 6 hours ago

---

The fastest way to improve an AI agent's code isn't a better prompt. It's better guardrails.

Greg Pstrucha from Sentry joins us to break down "Stop Prompting," his approach to raising the floor on agent output with deterministic tooling: linters, strong tests, and a real type system. Spend 30 minutes wiring these up and you stop wasting time correcting the same elementary mistakes on every review.

We get into:

• Why deterministic rules beat repetitive prompting for agent quality
• Fewer, stronger tests over more tests, and LLM-as-judge for evals
• How Sentry evaluates Seer, its debugging agent, all the way to the PR it produces
• Why "all agents are coding agents," even when they're just fetching telemetry
• The snake oil of quantifying quality: cyclomatic complexity and test coverage as metrics agents just game
• What it would actually take to stop reviewing agent code by hand and go straight to main

To learn more about Sentry, visit sentry.io/. To learn more about WorkOS, visit https://workos.com.