Mastra published the real numbers behind its AI agent claims

Mastra shipped Factory into beta on September 8 and took Product Hunt's top spot September 9 with 468 upvotes. The headline claim is that AI agents author 25-35% of merged PRs and close 50-60% of issues.

Most vendor posts stop at the range and let the top number do the marketing. Mastra didn't. The same post gives audited counts that fall well short of its own claim: agents authored 17.0% of merged PRs, 277 of 1,627, and closed 28.5% of issues, 222 of 778. Printing the figure that undercuts your own headline, in the post making the headline, is rare enough to note on its own.

The more interesting change is structural. In July, Factory ran fully automatic end to end. Backlog imports then strained infrastructure and produced work nobody could review. Mastra responded by splitting the work into six stages, Intake, Triage, Planning, Build, Review, Done, each one switchable between manual and auto independently.

A system producing PRs faster than humans can review them is a queue problem, not a scaling win. The redesign is Mastra's admission of that, not a feature announcement dressed up as one.

What the post doesn't say is whether the 25-35% and 50-60% range was set before the redesign or derived from the same audit as the lower numbers. If it's a pre-redesign target left in the post after the real figures came in under it, that's a different story than a range meant to describe current performance, and Mastra's post doesn't say which one this is.