Essays on AI-assisted software development: harnesses, mechanical gates, review at agent throughput, and keeping control of codebases AI writes most of.https://prasunsrivastava.com/en-ushttps://prasunsrivastava.com/writing/run-every-eval-more-than-once/https://prasunsrivastava.com/writing/run-every-eval-more-than-once/I ran the same AI code review five times at temperature zero, with everything held identical, and got a different answer two times out of five. A single run can push you to deploy the wrong model without ever knowing. This is how I size the number of eval repeats by what the decision is worth.Wed, 22 Jul 2026 00:00:00 GMTprasun@prasunsrivastava.com (Prasun Srivastava)https://prasunsrivastava.com/writing/review-gate-vs-single-prompt/https://prasunsrivastava.com/writing/review-gate-vs-single-prompt/I rebuilt Cloudflare's multi-agent review gate and ran it against a plain single-prompt review, on 12 commits from a real codebase's own history. The single prompt caught more of the known bugs at about a quarter of the cost, and the runs surfaced 9 shipped bugs nobody had caught.Wed, 15 Jul 2026 00:00:00 GMTprasun@prasunsrivastava.com (Prasun Srivastava)https://prasunsrivastava.com/writing/auditing-my-own-ai-coding-harness-for-enforcement/https://prasunsrivastava.com/writing/auditing-my-own-ai-coding-harness-for-enforcement/I audited the AI coding harness I wrote for a production codebase in 2025: 12,000 lines of rules, zero mechanical checks. What I moved to gates, what stayed prose, and why agents made the enforcement layer necessary.Wed, 08 Jul 2026 00:00:00 GMTprasun@prasunsrivastava.com (Prasun Srivastava)