Open any engineering blog this year and you'll run into the same brag: some company merging an absurd number of pull requests a week, most of them written by agents instead of people. Stripe ships over 1,000 PRs a week through something they call "Minions." Shopify has a Slack bot named River that co-authors one in eight merged PRs across the whole company. Uber says agents are now attributed to more than 70% of its pull requests, and that number is still climbing. And a lot of these teams keep reaching for the same word to describe what they built: a software factory. I went down the rabbit hole on this one, because the idea is genuinely going to change what your day as an engineer looks like — whether you build one or not. What is a software factory, actually? Most of us still use AI like this: open Cursor or Claude Code, type "fix this bug," watch it work, read the diff, hit accept, repeat. You're still in the loop. Nothing happens unless you're sitting there driving it. A software factory is what happens when you replace the person driving with a pipeline. Roughly: An event comes in — a GitHub issue, a Sentry error, a Linear ticket, a Slack mention. An agent triages it — decides whether it's a bug, a feature request, or noise, and what to do next. An agent implements it — spins up a sandbox, reproduces the problem, writes a fix, runs the tests. An agent reviews it — reads the diff, scores the risk, flags anything scary. A pull request comes out the other end , and a human (or, in a few cases, another agent) merges it. Ticket in, specialized agents on the line, PR out. That's the factory. If you've ever used a Cursor cloud agent or GitHub's Copilot coding agent, you've technically already run a tiny one. Why this is happening right now The short version: AI made code cheap to generate, but it didn't make code cheap to read . Every maintainer of a popular repo is now drowning in issues and PRs that AI helped someone else produce in seconds. Vercel's own AI SDK repo is the case study everyone cites. By mid-2026 they were getting 100+ new issues a month, and had piled up over 1,000 open issues and nearly 800 open PRs. Their honest conclusion: no maintainer, however good, closes that gap by just working harder. So they built a factory — one agent each for classifying, analyzing, implementing, and reviewing — and it started closing the majority of incoming issues on its own, with a human still merging every change. That "human still merges every change" detail matters, and we'll come back to it. Who's actually running one A surprising number of companies have public numbers now, even if they don't all use the word "factory": Uber runs the biggest published example. As of its late-August 2026 engineering post, more than 70% of pull requests are attributed to local or cloud agents, weekly agentic requests grew 9.4x between February and August, and engineers have built over 3,600 reusable "agent skills" that fire 30,000+ times a day — covering code review, CI self-healing, on-call triage, and bug debugging. The more notable part: total AI spend has stayed roughly flat over that same stretch, because Uber has been aggressively optimizing cost per session rather than just letting usage run wild. Vercel built ai-sdk-factory in the open, so you can watch its PRs land in real time. Cloudflare + Astro built a triage pipeline that reproduces a bug, ships a preview build, waits for the original reporter to confirm the fix, and only then opens a PR. Astro's open issue count dropped from 200+ to around 30. It's open-sourced as triagebot-action . Warp launched "Warp Factories" in August, pitched as Terraform-for-agents — define the pipeline as code. They claim about 30% of their own internal tasks already run through it. Stripe ("Minions"), Ramp ("Inspect," now reportedly around 75% of merged PRs, up from 30% in January), Shopify ("River"), Spotify ("Honk"), and PostHog (whose review bot StampHog auto-approves roughly a fifth of PRs for about $300/month in tokens) are all running versions of the same pipeline under different branding. Factory.ai published a manifesto in June calling this "Factory 2.0" — an end-to-end, agent-native system engineers build once and then let build the rest of the software. BCG Platinion has gone as far as declaring an "Agentic Software Factory era," citing organizations reporting 3–5x productivity gains — a claim worth reading as a self-reported estimate, not an independently audited benchmark. Maybe the wildest data point: OpenAI has reportedly stood up an internal beta application — something like a million lines of code, tests, CI config, and docs included — built from an empty repository by three engineers who steered coding agents and never wrote the code by hand. Lights on vs. lights off The real argument in this space isn't whether factories work — it's whether a human needs to look at the output at all. Lights on : a human still merges the PR. This is what almost everyone above is doing. Lights off : an agent merges it too. No human ever reads the diff. StrongDM is the clearest lights-off example. Their internal rule is literally that code must not be written or reviewed by humans — engineers write extremely detailed specs, agents write and test the code against a simulated "digital twin" of services like Okta and Slack, and the loop runs until it converges. They open-sourced the tool that does this, called Attractor. Before you get tempted to try that on your own repo, it's worth watching Dex Horthy's HumanLayer talk on the subject. He ran a fully automated, lights-off factory for four months with zero human review, and it went badly — around month three, the codebase had rotted so much that a single bug took weeks to untangle. His point: coding models are trained to make tests pass, not to leave behind code that's still legible to a human (or another model) six months later. Nothing in a fully autonomous loop punishes bad design until it's too late. That's echoed by GitClear's independent analysis of 211 million changed lines of code, which found rising duplicate code blocks and higher short-term churn as AI-assisted coding scales — a reminder that a factory optimized purely for PR throughput can quietly manufacture its own maintenance debt. It's not proof AI caused every regression, but it's a real counterweight to the "70% of PRs!" headlines. So for now, almost every serious factory keeps the lights on for anything that matters. Why this should matter to you, specifically Even if this feels distant from your day-to-day right now, it's coming for the way engineering work is organized. A few shifts I think are already underway: 1. Reviewing code becomes the actual skill. Every team above says the same thing: once the factory opens dozens of PRs a day, review is the bottleneck, not implementation. Knowing what good code looks like — and catching what's subtly wrong — is a different skill from typing it out, and it's the one AI is worst at replacing. 2. You end up "improving the factory" instead of just shipping features. A bad agent run is a signal, not a failure: fix the prompt, add an eval, provision a missing credential. That's still engineering — it's just pointed at the system that does the work instead of the work itself. 3. You'll probably build a small one yourself. These aren't hard to stand up. If you want to try: Vercel's eve template — one-click deploy with four agents (classifier, analyzer, implementer, reviewer); label an issue factory and get a draft PR back, with merging staying human. Astro's triagebot-action — a label-driven state machine built on GitHub Actions; probably the clearest one to read if you want to understand how the pieces actually connect. Warp Factories — the "as code" version, currently in closed beta. Ramp's Inspect spec — the messier, more real-world version, sandbox snapshots and Slack classifier included. The one rule everyone agrees on, from Vercel to Uber to the smallest weekend project: start with one small, boring job. Get it reliable, measure it for a few weeks, then add the next station on the line. Warp thinks this ends up as standard as CI/CD within a few years — and given where the numbers already sit in September 2026, that doesn't sound like hype anymore. Are you running one yet? Or are you staying lights-on and skeptical for now?