How Factory is turning AI into ‘a junior developer in a box’

[ResidentialBusiness]

Many things remain uncertain about AI’s future impact on our lives. One that isn’t in doubt is that more and more of the world’s software will be written, at least in part, by software. Already, 25% of Google’s code is generated by AI, CEO Sundar Pichai said last October. By 2028, projects research firm Gartner, 75% of enterprise developers will use AI tools in their work.

This trend is reflected in programmers’ embrace of products such as GitHub Copilot and Cursor, which let them call on generative AI to fill in some of the specific code as they tackle a project—essentially a fancy form of autocomplete for software engineering. The next step beyond that is AI coding assistance that’s more agentic—capable of handling at least certain tasks from start to finish without constant supervision. That’s what a San Francisco startup called Factory offers with its platform, which officially debuts today.

“Our mission, at a high level, is to bring autonomy to software engineering,” says Factory CEO Matan Grinberg, who founded the company with CTO Eno Reyes. Its platform includes agents—Factory calls them Droids—that “automate tasks in the software development lifecycle, and in particular tasks that developers don’t want to do—testing, debugging, refactoring, migrations, all that ugly stuff,” explains Grinberg.

Factory aims to go beyond the code-autocomplete features offered by tools such as GitHub Copilot.

Rather than replacing something like the GitHub Copilot, Factory aims to provide a new layer for software projects that’s compatible with the processes and products a team already has in place. “For those who use Copilot, it’s like, ‘Okay, cool—I brought this down from [Factory],’” says Grinberg. “‘Now there’s some other things I want to do.’”

‘I had a bit of an existential crisis’

Grinberg and Reyes were respectively 24 and 23 when they started Factory in 2023—by all accounts, a whirlwind of an experience. It started when Grinberg, who’d studied string theory at Princeton and was working on a PhD in physics at Berkeley, suddenly began questioning his trajectory in life.

“I realized that I had spent basically the last decade obsessed with physics,” he says, “and I was only doing it because it was hard, not because I actually loved it. Which is obviously a horrible reason to pursue a career path. So, I had a bit of an existential crisis.”

Grinberg found a new obsession in AI, a technology whose destiny had been permanently altered by the arrival of ChatGPT. He was particularly fascinated by program synthesis (later better known as code generation): the science of teaching software to write software. He became a regular at AI hackathons, including the one where he met Reyes, who’d written his thesis on deep learning and worked on language models at Microsoft and Hugging Face. (The two had been at Princeton at the same time and—despite having overlapping circles of friends— had somehow managed not to know each other.)

At Hugging Face, “I was working with Databricks, Bloomberg, Grammarly—everyone wanted code gen,” says Reyes. He’d already collaborated on an AI code generator for finance applications. “There’s way more opportunity here,” he remembers thinking. 

Factory’s platform lets users select code and other data to incorporate into collections called Workflows.

This is where a classic Silicon Valley element of serendipity kicked in. At almost the same moment Grinberg and Reyes connected, Grinberg had sent an unsolicited email to Shaun Maguire, a partner at venture capital titan Sequoia, seeking career advice. “It’s very rare that cold emails actually turn into something,” says Maguire. But like Grinberg, Maguire had a background in high-energy physics theory. And when he learned that Grinberg had coauthored a paper with legendary physicist Juan Maldacena, the credential blew his socks off.

In person, Maguire was even more impressed by Grinberg: “I was shocked that not only does this guy like have the IQ to write a string theory paper with Juan Maldacena as an undergrad, but he also has charisma and sales ability and empathy.” He immediately encouraged Grinberg to pivot from PhD student to startup founder—a goal Grinberg now acknowledges he’d already had in mind, though he’d kept it to himself.

A week later, Grinberg returned with the idea that became Factory. By then, the GitHub Copilot had shown that AI could produce lines of code usable in a production environment. But as the name Copilot indicates, it was doing so under the watchful eye of a human programmer. Grinberg was thinking ahead to AI that could perform some straightforward tasks more independently—in other words, agentic AI, though nobody was bandying around the term at the time.

“His point was, ‘Copilot is great, but soon we’re going to have junior developers in a box,’” remembers Maguire. “And that’s what he wanted to work towards.” Sequoia helped get Grinberg and Reyes on their way by leading Factory’s $5 million seed round of funding. It followed up by leading a $15 million Series A round in 2024.

‘Delegating away some tasks’

Grinberg emphasizes that Factory doesn’t envision AI taking over coding in a sweeping fashion anytime soon. Instead, the exact nature of the human-computer collaboration will vary from area to area. “Software developers of the future will be delegating away some tasks,” he says. “They will be pairing with AI on others. And they will be more directly hands-on working on some things with suggestions from AI as they do it.”

In a demo, he showed me how the company’s platform uses AI to let engineering teams create shareable, easily digestible collections of all the code and other data associated with a particular project. Known as Workflows, they don’t just help humans keep tabs on what’s where. They’re also a starting point for the Droid agents, which can take on useful grunt work relating to the code encompassed by a Workflow.

For example, many programmers use a system called Mermaid to create diagrams about works in progress—say, a chart documenting all the dependencies that various blocks of code might have on each other. Instead of personally writing the lines of JavaScript necessary to create a Mermaid diagram, a user might have a Droid do the job and then save the results as a code snippet for later use. Similarly, if a project has code that lacks comments—embedded explanations documenting what the software is doing and how it does it—a Droid can add them.

Among the software development tasks Factory’s AI can assist with is one of the most basic of them all: debugging.

Those are examples of tasks that many teams might happily offload to AI. But the whole point is that users can ask Droids to undertake assignments on the fly—“something as spontaneous as ’Hey, can you generate me a customer usage dashboard?’” says Grinberg. Such a dashboard could draw on any relevant data the user added to the Workflow in question. And like everything Droids create, it would be reusable.

Of course, there’s nothing new about using dashboards to help wrangle complex projects. In the past, however, they’ve been one-size-fits-all tools hard-coded by a platform provider. By using generative AI, Factory wants to pioneer a more fluid approach in which its customers can call on Droids to construct the functionality they want when they want it. Instead of being “set in stone,” says Reyes, a team’s working environment can consist in part of “LLM-generated, malleable pieces of content.”

That malleability includes the ability to choose the large language models that power Factory’s AI: “We support everything, basically,” says Grinberg. Even freshly-minted, cutting-edge models are on the list, including xAI’s Grok 3 (announced last week) and Anthropic’s Claude Sonnet 3.7 (announced on Monday).

By riding atop all the major LLMs, Factory expects to get more powerful as they do, opening up scenarios where Droids grow competent at work that’s presently beyond their skill set. “Building for what’s going to be possible in a year or two is how we got here first,” says Reyes. “And we’re already building for what will be possible a year or two in the future.”

Which is not to say that Factory fully understands how its customers will use its platform. As with all things generative AI, nobody knows for sure what it can and can’t do well until someone gives it a try.

“We just want a ton of people to see this new paradigm of interacting with software,” says Grinberg. “It’s just going to be so much fun to have all these people who have really strong opinions get their hands dirty.”

View the full article