Who Reviews Agent-Written Code? The Capacity Problem

Coverage of Cursor's Compile demo carried one figure that did more work than the rest of the launch combined: a slide claiming Origin handled 22.6 commits per second against a single repository. Treat the number the way this site treats every pre-launch figure, as a demo claim rather than a benchmark. The reaction it produced is the interesting part, because even at a hundredth of that pace, no team on earth reads every change.

Human review capacity is measured in diffs per day, not per second. A careful reviewer reconstructs intent, checks the change against what was asked, and decides whether to trust it, and none of those steps got faster because a model wrote the code. So the queue moves. Writing stops being the constraint and everything after it, review, merge, CI, becomes the place work piles up.

That mismatch is not a flaw someone forgot to fix. It is the design problem this whole product category exists to answer, and it is worth being suspicious of anyone who claims it is already solved.

Origin's launch thread went up on June 16, 2026, pointing at a page that was a tagline and a signup form. The thread was small but the objections were the right ones, and they have aged well precisely because Cursor still hasn't answered them.

The middle objection is the sharp one. Storage and merge throughput are engineering problems with known shapes; review capacity is a human limit. Speeding up how fast machines write and land code does nothing for you if nobody can vouch for what landed. Any honest answer to the thread has to describe who, or what, is doing the vouching.

Three mechanisms carry the load right now, and none of them requires Origin. Each attacks a different part of the capacity problem, and each has a failure mode worth naming up front.

They compose. Stacks shrink the unit of review, bots clear the mechanical findings, and the queue makes landing safe, which leaves human attention for the one thing none of the three can do: deciding whether the change matches the intent. That composition, not any single tool, is the current answer to the capacity problem.

All three mechanisms have shipped for years at the company that now builds Origin. Cursor acquired Graphite in December 2025, and Graphite's products map onto the table above almost one to one: stacked pull requests are its signature workflow, its merge queue promises to keep your main branch green with CI optimizations that skip redundant runs, and Diamond is its AI reviewer. Tomas Reimers, a Graphite co-founder, gave the Origin demo at Compile.

That lineage is the strongest evidence about what Origin's review layer will look like, stronger than anything on the waitlist page. What remains unconfirmed is the integration: whether Origin bakes stacks, queue and AI review natively into the host, and what that changes versus running Graphite's tools against GitHub the way teams do today. The review guide tracks what coverage has reported there.

The honest ceiling: tooling changes what humans review, not whether they must. The skill shifts from reading every line to judging a stream of triaged changes and knowing which ten of the hundred need a careful human eye. That is a real skill, it is trainable, and it is still a finite budget of attention.

Three limits survive every tool in the table above.

• Bot approval is not accountability. When a change breaks production, the person who merged it owns it, and no first-pass reviewer changes that.
• Intent-checking can't be delegated. Only the person who asked for the change knows whether the diff does the right thing, as opposed to a passing thing.
• Volume pressure breeds rubber stamps. A queue of green-checked diffs invites approve-on-vibes, and the failure is invisible until something merged that nobody understood.