Priya Subramaniam

Version code, prompts, models, and tools as a single release unit.

Confusing these layers wastes engineering budget on the wrong problems.

Parallel agents need isolated storage to avoid colliding over shared code.

Escalation design matters more than confidence scores for autonomous agents.

Agents produce too many PRs for humans to catch flaky tests anymore.

Old velocity metrics fail as AI agents flood pipelines with code that doesn't ship.