Software factories and code reviews with Harjot Gill
Harjot argues that AI agents and software factories make human code review obsolete, justify enormous token and compute spend, and reward high-agency, strategy-game-minded engineers, though the host remains skeptical on several points.
Watch on YouTube ↗Charlie's summary
- ▶ 15:48
Software factories: roughly 2 of 20 chats are hands-on crafting, the other 18 are agents hill climbing against a verifier
Harjot splits AI use into the potter's wheel, shaping software one chat at a time, and the factory, where agents loop on a goal. Hill climbing works when an agent can verify its work. CodeRabbit uses TLA+, simulations, Lean theorems and API emulators. One agent ran six weeks building a Rust Git write-ahead log on object stores, and Harjot read none of the code. Each release is a “turn in a strategy game” the agent learns from.
- ▶ 18:50
Small teams now need Facebook-grade release tooling, because 30 to 40 engineers can ship like thousands
Harjot says staged rollouts, merge queues and testing in production used to be only for giants like Facebook and Netflix. With agents, a 40-engineer team ships at that volume and needs the same risk management. He sees an opening for engineers from those companies to build products for it. CodeRabbit ships every 30 to 45 minutes, using feature flags and progressive rollouts that agents watch.
- ▶ 25:25
Human code review is dead, but human review isn't: the pull request becomes the gate for judgment and risk
Harjot thinks that agents, designers and alerts now flood in code with no upstream planning. So the PR is where a human decides whether it is worth shipping and what the risk is, since “human taste is the bottleneck now.” At CodeRabbit the unit of change is a release. Engineers review risk levels in Slack rather than reading code. Some PRs go straight to production, even from interns, and rollback takes 10 to 15 minutes.
- ▶ 30:32
A driving license for engineers: ship too many regressions and you lose the right to skip review
CodeRabbit ties each production fix back to the PR that caused it and measures how long the regression lasted. Developers on a review-exempt list who are careless get strikes. Hard prompting with no follow-through can mean 100 hours of exposure. After enough strikes the license is revoked and they return to human review until they prove themselves. Harjot says some people steer agents diligently while others are reckless, and the license penalizes the second kind. The host liked this idea best.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.