Determinisim In The Agentic World

When agents can write code faster than you can review it, quality becomes the bottleneck. This talk covers how to maintain high confidence in agent generated code by applying Zed’s three engineering pillars alongside a fourth rule: always give the agent a deterministic feedback loop. Using Zed’s git graph as a case study, I’ll walk through how property based testing lets an agent verify its own changes against invariants instead of relying on human review to catch regressions. I’ll also preview a debugger tool call that gives agents direct access to runtime state. The core idea is practical. If an agent can’t prove its output is correct, the velocity it gives you is a liability.