Houck: the invisible output of review is the org
DX’s Brian Houck on Meta’s review crush under agents — RADAR’s risk-stratified automation works, but automating diffs without protecting knowledge transfer and intent debt just hollows the team.
Source: What are code reviews even for?
Brian Houck (DX / Microsoft Distinguished Scientist) starts from the queue math everyone feels. At Meta, significant lines of code per human-landed diff up 106% in a year, diffs per developer per month up 51%, more than 80% of that growth from agentic AI — while 24-hour review coverage falls and some groups stare at thousands of pending reviews. DX’s own data: median PR size up 64%. Developers already want review to be ~7% of their time. Asking them to review more AI output is not a capacity plan.
The useful move is the question before the tool pitch: what was review solving before AI? If it were only defect finding, full automation looks inevitable. Bacchelli & Bird already showed defect-related comments were only ~14% of review comments; most of the value was alternative solutions, knowledge transfer, and awareness of teammates’ work. Houck’s line: “The visible output of code review is better code. The invisible output is a better engineering organization.”
He does not reject automation. Meta’s RADAR is the receipt he leans on — risk-aware auto-review for a selected low-to-medium slice, humans on the rest: 535k+ diffs reviewed, 331k+ landed, revert rate ~⅓ of non-RADAR, incident rate ~1/50th, 3.3× faster median close. The design claim is eligibility and risk stratification, not “the model says LGTM.” Developers in DX’s research want AI to catch security/compliance, flag high-risk changes, and clear routine — and explicitly do not want AI that auto-merges or takes final accountability.
The warning is slower than a bad merge. Automate the diff and you may have reviewed that diff; you have not transferred knowledge, built shared ownership, or given a junior a window into how a senior reasons. Storey’s framing shows up as cognitive and intent debt — the gap between what the system does and what the org collectively understands about why. It arrives as a series of reasonable skips: this one’s low risk, this approval can be automated. Measure review health beyond throughput, or you will only notice when an outage needs the conversation that never happened.
Why it matters for code review
Software factories that treat review as a defect firewall will automate the ceremony and lose the org. The control plane Houck is arguing for is the same one good agentic SDLCs need: risk tiers, conservative eligibility, auditability, humans still named on acceptance — plus deliberate measurement of whether juniors still learn and whether architectural understanding still spreads. Speed the routine; protect the conversations that make the team able to own the system the agents are flooding.