The part I'd underline is the scaling effect: once Codex can run dozens of tasks in parallel, test coverage stops being a background quality issue and becomes the main control surface.
A weak suite does not just miss one bad change; it lets many plausible-but-wrong changes arrive at review at once. The harder engineering problem may be less "can the agent write code?" and more "can our verification stack tell correct code from merely green code?"
Exactly, parallelism turns a weak suite from a slow leak into a flood of plausible-but-wrong PRs. The real bottleneck moves from writing code to verifying it, and that layer barely exists yet.
Paolo, excellent breakdown.
The part I'd underline is the scaling effect: once Codex can run dozens of tasks in parallel, test coverage stops being a background quality issue and becomes the main control surface.
A weak suite does not just miss one bad change; it lets many plausible-but-wrong changes arrive at review at once. The harder engineering problem may be less "can the agent write code?" and more "can our verification stack tell correct code from merely green code?"
Exactly, parallelism turns a weak suite from a slow leak into a flood of plausible-but-wrong PRs. The real bottleneck moves from writing code to verifying it, and that layer barely exists yet.