How StrongDM’s AI team build serious software without even looking at the code

StrongDM's AI team ships production code written entirely by agents. No human code review. They pinpoint the inflection to Claude 3.5 Sonnet (October 2024), when long-horizon agent workflows started producing correct code instead of compounding errors. Validation uses "scenarios" -- end-to-end user stories kept outside the codebase as holdout sets, like model training validation data. Success is measured probabilistically ("satisfaction" scores) rather than pass/fail. Basically aggressive external QA, applied to agent output.

Visit Original Article →