The Pragmatic Engineer newsletter ran a reader essay challenge, and the winning entry, published on November 3, is a first-hand account of what AI coding looks like at full speed. Its author, Neil Kakkar, has nine years of experience, including Bloomberg and PostHog, and is now the first engineer at Tano, a startup of about ten people.
The numbers
At earlier jobs Kakkar merged 8 to 20 pull requests a month. At Tano he ships more than 200, about 11 times his PostHog output. He reports a revert rate roughly six times lower, mostly because problems are fixed forward, and a median time from opening a PR to merging it of under an hour. The whole team merges about 30 PRs a day.
The setup
Everything runs on Claude Code. Each engineer has one Claude Max 20x subscription; Kakkar estimates his usage would cost more than $8,000 a month at API prices. He works in the Claude Code desktop app with its built-in browser rather than the command line, keeps 3 to 5 tasks going in parallel and lets agents run for one to two hours per feature, including their own checks and screenshots. Every PR gets an automatic review from a Claude Code agent in GitHub Actions, with a deeper review agent for risky changes. On most PRs, that agent is the only reviewer. Context about each product area is written down in AGENTS.md files so the agents can find it.
What went wrong
He lists the failures plainly. An agent multiplied currency amounts by 100, which is wrong for currencies such as the Bahraini dinar; it was caught after deploy and fixed within the hour. A blocking database migration on a large table stopped data ingestion. Another migration without safeguards paid some creators twice. Rules like these now live in the review agent's instructions.
The human costs are on the list too: constant context switching, fewer people who know any given piece of code, generic-looking design, and hitting plan limits when agents run in loops. Kakkar's own rule is to stop at 6 pm and leave his work devices at the office.
The editor's caveat
Gergely Orosz, who runs the newsletter, adds that across the industry PR output has roughly doubled since January, so this case is far above average, and that speed alone does not prove a product is succeeding. Mature companies, he argues, should change their processes more carefully.
What this means if you are choosing a coding agent
The gain came from the process, not just the tool. Written-down context, an automatic reviewer on every PR and rules learned from incidents did as much as the model.
Budget for limits. Even the top Claude plan ran out when agents worked in loops, so check the allowance before you plan around overnight runs. If you are comparing options, see Codex vs Claude Code and our list of AI tools for coding.