Marcus: The team behind Next.js, the web framework, was looking at a backlog of over 3,000 open bug reports on GitHub.
Sierra: Three thousand. That's not a backlog, that's a declaration of bankruptcy.
Marcus: It's a mountain. And it peaked in January 2025. Even by summer of this year, they were still at over two thousand.
Sierra: So they just set up a bot to start closing tickets on people. I feel like I've been on the receiving end of that. You file a detailed bug report and a robot just says 'no' and locks the thread. It's infuriating.
Marcus: That's what I thought this was going to be, but it's not. They actually tried that first — just a simple script. If an issue had no activity for 18 months, a bot would mark it stale and then close it.
Sierra: And I'm guessing that didn't go well.
Marcus: It was a terrible signal. A timestamp can't tell you if a bug is a real, neglected problem or if it's already been fixed and just not closed. They were closing reports they should have kept. So they killed that idea.
Sierra: Okay, so what's plan B? Hire a hundred interns?
Marcus: Basically, yeah. But the interns are AI. They built an agent they call `closability`.
Sierra: Closability. Love the name.
Marcus: And its job is not to close anything. Its job is just to do the research. It spins up a little sandbox environment, reads the entire GitHub issue conversation, and then goes hunting. It searches for related pull requests, commits, documentation... it even tries to reproduce the bug on the version it was reported on, and then again on the latest version.
Sierra: So it's an AI detective, not an AI judge.
Marcus: Exactly. And this is the key part: the agent is read-only. It can't comment, it can't push code, it can't close the issue. All it does is produce a report with a confidence score saying, 'Based on my research, here's why I think this issue could be closed.'
Sierra: But how does it decide? The article says it has to weigh 'incomplete and sometimes conflicting evidence'. That's the part that feels like hand-waving. How does it know a user's failed reproduction isn't more important than its own successful one?
Marcus: I'll be honest, the post doesn't really detail the weighting algorithm. It just says they kept the confidence score 'conservative'. That feels like the secret sauce they're not fully explaining. But, the process itself is what's interesting. The agent's report goes into a queue for a human maintainer to look at.
Sierra: So a human makes the final call.
Marcus: A human makes the final call. And they ran this on the entire backlog. Two hundred agents running at once, each taking about 30 minutes per issue. And in three weeks, they closed 1,462 issues.
Sierra: Okay, so what was in that pile of 1,400 issues? Was it all just trivial stuff?
Marcus: This is the best part. The number one reason for closing, at 37 percent, was 'Already fixed'. The fix was in the codebase, but no one ever went back to close the ticket.
Sierra: So the problem wasn't the bugs. It was the bookkeeping.
Marcus: Nineteen percent were duplicates. Another 16 percent were 'expected behavior'. It wasn't a bug, it was a feature. All this stuff was just noise burying the real problems. And here’s the kicker: they built in a little action where anyone can comment 'Reopen' on a closed issue and it will automatically reopen.
Sierra: The escape hatch. How many people used it?
Marcus: Three.
Sierra: Wait, three? Out of almost fifteen hundred?
Marcus: Three. A 99.8% accuracy rate. My skepticism kind of melted when I read that.
Sierra: Alright, that is... genuinely impressive. So they cleaned up the backlog. But now what?
Marcus: Well, they're not stopping. The blog post says they're now letting the agents start to auto-close issues. Small scale. Up to 25 a week.
Sierra: Ah, there it is. I knew the human-in-the-loop was too good to be true. So we're back to the robot that just says no.
Marcus: It's a little more nuanced. It only does it for issues with a super high confidence score, and a second agent has to independently review it and agree. And the reopen button is still there.
Sierra: But that feels like the start of a slippery slope. What happens when a tricky, intermittent bug gets auto-closed because the AI couldn't reproduce it in its clean sandbox environment? That person just gets their issue ignored by a machine.
Marcus: But what's the alternative? That same tricky bug sits in a 3,000-issue backlog and gets ignored by a human because they never even see it. At least this way, the real, actionable bugs have a chance of floating to the top. This isn't about replacing the developers, it's about giving them a team of robot assistants to handle the paperwork.
Sierra: I suppose if the choice is 'maybe get closed by a robot' versus 'definitely get buried by an avalanche,' I'd take my chances with the robot.
Marcus: That's how they frame it. As AI speeds up development, it also creates more issues to triage. You either drown, or you build smarter tools to stay afloat. I'm Marcus.
Sierra: And I'm Sierra.