Anthropic's Cloud Tasks Point to 'Software Factory' Future

Anthropic quietly shipped a feature last week that developer Ray Amjad thinks reveals where AI-assisted coding is actually heading. It's not the feature itself that matters—it's what Anthropic is building infrastructure for.

The update lets Claude Code run scheduled tasks on Anthropic's servers instead of your local machine. You give it a GitHub repo, a prompt, and a schedule (hourly, daily, whatever), then close your laptop. The AI runs in the background, churning through your backlog.

Amjad demonstrated setting up what he calls a "Sentry fixer"—a task that checks his error tracking system every hour, identifies bugs affecting three or more users, writes fixes, and opens pull requests. While recording his demo, the system actually found and fixed a bug, then sent him a Telegram message explaining what it did.

That's the headline feature. Here's what it means.

The Factory Floor Reconfiguration

The concept Amjad keeps returning to is the "software factory"—a term floating around developer Twitter that describes AI agents working around the clock on your codebase. Bug reports come in through Gmail, get triaged automatically, turn into fixes, generate PRs. Feature requests from user feedback boards get prototyped overnight. CI failures get investigated and resolved. Documentation updates itself.

"We're increasingly moving towards a software factory where you can kind of imagine that overnight an AI assistant is basically going through your GitHub issue backlog tackling each feature and then making brand new PR for you to review when you wake up in the morning," Amjad explains.

Connect enough services—Sentry for errors, Linear for tickets, Slack for alerts, user feedback platforms—and you've got an assembly line. Information goes in one end (bug report, feature spec, customer complaint). Code comes out the other. Humans approve or reject.

Amjad's framing is revealing: "The average developer role is slowly shifting from being a worker along that assembly line to being like a factory floor manager where agents are doing most of the production and you're just like testing the output, approving it, like declining things and then just optimizing the factory line."

Factory floor manager. That's the role being built for here.

Building for Tomorrow's Models, Not Today's

Here's where it gets interesting—and where Amjad is honest about the current gap between vision and reality.

If Claude Code already introduces bugs into your codebase (and it does), why would you let it run unsupervised overnight, making more commits, opening more PRs?

Amjad's answer comes from Anthropic's own playbook: "They're building for where the models will be 6 months from now rather than where the models are right now."

Right now, with Opus 4.5 or 4.6, you're probably reviewing everything the AI writes. With Opus 5, maybe just a quick glance. Eventually—the vision goes—you get to "lights out" operation where "errors are automatically discovered, bug requests and features are automatically identified and then humans only have to put in a small amount of judgment."

This is infrastructure speculation. Anthropic is laying pipes for water that hasn't arrived yet. The feature itself is modestly useful today. But if models improve on the trajectory everyone expects, the architecture is ready.

And there's a business logic to it. As Amjad notes, this model "essentially means that we will be consuming even more tokens from them." An AI that runs 24/7 burns through a lot more API calls than one you use interactively for a few hours.

What Gets Lost in Translation

What Amjad doesn't explore much—though he gestures at it with his half-joking farm comment—is what happens to the actual experience of software development.

If you spend your day writing specifications that get built overnight, what are you actually doing? Product management? Systems design? Quality control? It's unclear, partly because we're describing a job that doesn't quite exist yet.

There's also the question of what "oversight" means when you're reviewing PRs generated by something that works faster than you can keep up with. Amjad envisions agents that might send you voice messages when they hit ambiguity, asking for direction. That sounds helpful until you imagine fielding dozens of these decision points daily while also trying to... manage the factory?

The historical precedent here isn't encouraging. When humans become "floor managers" for automated systems, we're terrible at it. We zone out during monitoring tasks. We overtrust systems that are right 95% of the time until they catastrophically fail. We lose the tacit knowledge that comes from doing the work ourselves.

Amjad himself seems torn. He describes developers who want to "quit my job, open a cafe, move to the countryside, and start a farm" and adds: "I don't really blame you because I also feel like that sometimes."

The Actual Product Being Built

Pull back from the factory metaphor and what Anthropic has actually shipped is more modest: scheduled remote tasks with connector integrations. It's useful. It's iterative. It makes Claude Code stickier.

But the framing around it—from Anthropic and from observers like Amjad—is about something larger. The product roadmap here isn't just features. It's a particular vision of what software development becomes.

That vision might be right. Models might get good enough that autonomous overnight coding makes sense. The factory might work.

Or we might discover that the metaphor itself was wrong—that software development isn't actually a factory process that benefits from 24/7 operation, that the human judgment currently required at every step isn't a bug but a feature, that speed isn't the main constraint.

Either way, Anthropic is betting you'll want the infrastructure in place when you decide. The question is whether developers are choosing this future or just preparing for it because it seems inevitable.

— Marcus Chen-Ramirez, Senior Technology Correspondent