When the machines write the code, who reads it

Two threads from the same week point at the same wall. A security founder said out loud what a lot of teams are quietly assuming: within a year, most code that ships will be reviewed by AI, not a person, simply because review is now the bottleneck and nobody will tolerate it for long. Days later, Anthropic shipped review levels for its coding tool, where at the top setting a fleet of reviewer agents reproduces each finding independently before it counts.

So the shape of the near future is clear enough. Agents write the code, and other agents read it. The humans move to the edges.

I want to grant the strongest case for this, because it is real. When agents open dozens of pull requests a day, no human is reading them line by line, and pretending otherwise is worse than admitting it. A reviewer that never gets bored and checks every diff beats a tired person skimming the twentieth one.

Having said that, an AI reviewer is worth exactly what it catches, and the vault’s eval notes are blunt about the failure mode: a check can reward the appearance of correctness while missing the thing that actually matters. A reviewer that flags style and misses the authorization bug is not review, it is theater with good production values. The whole reason review exists is that someone can be held to what they approved.

So I do not mind machines reading the code. I mind “an AI reviewed it” becoming a sentence that ends the conversation. It should start one. Reviewed by what, at what effort, checking for which failures, and can you see the receipt. A rubber stamp does not get more trustworthy because it stamps faster.