Vercel's AI SDK factory pins trust to an agent configuration rather than a contributor

Two numbers in Vercel's write-up measure different things, and only one of them looks like a backlog. Authoring 25 to 35 percent of merged PRs is a production share. Closing 70 to 80 percent of issues is a disposal rate ^[12]. Apply the disposal rate to the late-June pile of more than a thousand open issues and you get 700 to 800 ^[21], but the reported figure is a rate four weeks into operation, and inflow did not stop while the agents ran.

For that rate to transfer, the pipeline has to survive its first stage. The pattern runs triage, then reproduction, then a fix or feature, then a review, then a hand-back to a human to merge ^[3], with separate agents for reproducing the bug, applying the fix and reviewing it ^[7]. Everything downstream of reproduction depends on a failure a sandbox can produce on demand. AI SDK is a package pulling more than 20 million npm downloads a week ^[5], so a large share of its reports are one wrong return value in a small script. If your issues are API design arguments, or races that only show up on one user's hardware, the reproduce agent stalls and the rest of the line has nothing to work on.

The mechanism worth copying is where trust lives. Vercel engineer Lars Grammel described it as trust in a particular agent configuration: a specific agent with a specific optimised prompt that history shows was successful on a certain category of bugs ^[9]. He advised open source projects to run their own agents and their own setup rather than trusting the community, on the grounds that it cuts review time ^[10]. That is the whole argument for closing the door. An outside PR may come from a better model, but it arrives with no measured record attached, and Vercel says it trusts its own agents over community-run ones ^[8]. Flue's contributor guide names what it is filtering, calling them "Drive-by AI slop PRs" ^[18].

Astro, at 62,000 GitHub stars, adopted the same idea ^[13]. Creator Fred Schott told Latent Space it has totally shifted in the last six months, with automations handling triage, reproduction, and getting the user to verify the bot's suggested fix before a maintainer looks at it ^[15]. That last step is the real saving: reviewer attention arrives at the end of a chain where the reporter has already confirmed the patch. Schott says the effect is that issues became a thing the team prioritises every week rather than a backlog it constantly trims, which he had not seen in a decade-plus of open source ^[16]. The auto-triage work is what led him to build Flue ^[17], whose guide pairs the team's own expertise with the best available state-of-the-art LLMs it can reach ^[20].

If I were adopting this, I would ship the reproduce agent alone and measure it for a month, because its output is the only one a human can check in ten seconds: the failing case runs, or it does not. The rest of the factory is worth building once you can name the bug categories your configuration has a record on. Pull requests have been open by default since GitHub invented them 18 years ago ^[1]; the projects closing them, tldraw among them, are the ones that already automated the other side of the door ^[2].