VERIFICATION-LED AI CODING: A CREDIBLE 1M-LOC MILESTONE
Paul Dix argues that AI, steered by a strong verification oracle, produced and hardened a million-line production codebase. In a quote surfaced by [Simon Willi...
Paul Dix argues that AI, steered by a strong verification oracle, produced and hardened a million-line production codebase.
In a quote surfaced by Simon Willison’s Weblog, Paul Dix claims AI generated ~1M lines of code and refined it to reliable software at massive scale.
His point: with the right verification harness and clear direction, agents can iteratively build complex systems and “grind” them into working production code.
If agents are guided by strong oracles, big rewrites and migrations become a specs-and-tests problem, not a headcount problem.
Quality shifts left: invest in behavior locks (contracts, goldens) and let agents iterate until they pass.
-
terminal
Pilot an agentic port of a small service using a golden dataset and contract tests; track defect rate, review churn, and perf deltas.
-
terminal
Build a verification loop (contract tests, fuzzing, static checks, differential testing) and gate merges on pass rate and SLO budgets.
Legacy codebase integration strategies...
- 01.
Target components with stable behavior and good observability; freeze interfaces and record goldens before any AI rewrite.
- 02.
Roll out behind flags and canaries; enforce style/lint rules to catch shallow patterns and duplicate logic.
Fresh architecture paradigms...
- 01.
Go spec-first: define contracts, acceptance tests, and performance budgets, then let agents scaffold to the tests.
- 02.
Capture an oracle early from real traces plus synthetic edge cases to constrain behavior.
Get daily AGENTIC-WORKFLOWS + SDLC updates.
- Practical tactics you can ship tomorrow
- Tooling, workflows, and architecture notes
- One short email each weekday