A San Francisco company arguing that software engineering is mostly delegation in disguise - and that the colleagues you delegate to do not have to be human.
It is a Tuesday morning at a large bank. An engineer drags a ticket from her Jira board into a browser window. She types two sentences of context, hits send, and walks to get coffee. Before she finishes the cup, a pull request is sitting in her inbox: tests written, types updated, an explanation of the trade-offs in plain English. The author of the PR is not on the engineering org chart. It is a Factory Droid, and there are now thousands of them on payroll across enterprises that pay Factory by the seat.
Factory is the three-year-old company that built those Droids. It calls what it does agent-native software development - a phrase that sounded like a research slogan in 2023 and now sits inside earnings calls. Its pitch is unfashionably blunt: the AI coding category will not be won by smarter autocomplete. It will be won by something closer to a colleague.
The company spent its first two years quietly fitting that habit to the people who are hardest to surprise - principal engineers at banks, insurance carriers, large software vendors. In April 2026 it raised $150M at a $1.5B valuation, and the quiet years ended.
Walk into any company with more than a hundred engineers and the dirty secret is the same. The thrill of greenfield code is a fraction of the day. The rest is upgrading a deprecated framework, hunting down a flaky test, reading a Slack thread to figure out what a PM actually meant, getting paged at 2 a.m. for a service nobody has touched since 2021. Engineers were trained to design systems. They spend their time being archaeologists.
The first wave of AI coding tools, charming as it was, did not really address this. Autocomplete in an editor is wonderful when you are already writing. It is useless when the question is which file should I be in? Or worse - which repo, in which monorepo, in which version?
Factory's founders looked at that gap and made an awkward observation. The work that consumes most of an engineering organization - ticket triage, dependency upgrades, incident response, doc rot - is structured enough for an agent and unstructured enough that no SaaS tool has ever really cracked it. So they decided to crack it.
Matan Grinberg was a few years into a UC Berkeley physics PhD when he met Eno Reyes at a LangChain hackathon. The thing they built that weekend - a small system that could solve coding tasks end to end without human babysitting - was the kind of demo most people show their friends and then forget. Grinberg did not forget. He cold-emailed Shaun Maguire at Sequoia, who told him to stop wasting time on his thesis. He did.
Factory was incorporated in 2023. The first product was a single Droid that could take a GitHub issue and open a pull request. It was rough. It was also, in customer demos, eerie.
Grinberg's note to Sequoia was, by his own telling, embarrassingly short. He explained the hackathon prototype. Maguire told him to drop out. Three years later, Sequoia is still on the cap table - alongside Khosla, NEA, NVIDIA, J.P. Morgan, Blackstone and Insight.
The bet underneath all of this is unfashionably specific. The founders did not bet that AI would replace engineers. They bet that AI would replace the parts of engineering that engineers already hate. The parts that fill the time between thinking.
Factory ships a small fleet of agents, not one general assistant. Each Droid has its own cognitive architecture, its own tools, and its own model routing. The marketing department calls them Droids and means it.
Picks up a ticket. Reads the codebase. Writes the change. Opens a pull request and waits politely for review.
Your on-call partner that does not sleep. Triages the alert, finds the root cause, writes the post-mortem doc.
Turns a rambling Slack thread into a coherent spec. Grooms the backlog. Routes the work to humans or other Droids.
Reads the ADRs nobody reads. Answers the question with citations, so the new hire stops asking the same one.
The platform lives in a browser. That choice is not aesthetic. The IDE assumes one human, one machine, one file at a time. Factory assumes one human steering many agents, possibly in many repos, possibly across many branches. The browser turned out to be the only place that geometry fit.
Most AI coding companies show off lines generated. Factory shows off hours not spent. Across its customer base the company puts the cumulative figure at roughly 550,000 engineering hours - an average of about 2,300 hours per organization, which it translates to a 20% reduction in development cycle time. Take the percentage with the usual grain of salt. Take the trend seriously.
Across Seed, Series A, B and C in roughly thirty months.
Post-money, April 2026, led by Khosla Ventures.
Cumulative engineering hours, according to Factory's own number.
Average reduction reported across customer orgs.
The customer roster Factory does talk about leans enterprise - financial services, large software vendors, regulated industries. The CEO of MongoDB has publicly cheered them on. Anthropic has run a customer story on them. The list of names Factory will not yet talk about is, by all accounts, the more interesting one.
Factory's stated mission is to bring autonomy to the software development lifecycle - which, stripped of the deck-speak, means letting an engineer plan a change in five minutes and watch four Droids ship it. The point is not to remove the human. The point is to put the human further up the stack, where the interesting decisions actually live.
That is also where the company's reading of the next few years lives. Grinberg has been candid that the hardest problems in agents are not knowledge problems but reliability problems. Long-horizon agents are wrong in subtle ways. Enterprise customers will not tolerate subtle wrongness. So Factory has spent unfashionable amounts of time on things like context management, evaluation harnesses, and governance - the parts of agent design that do not make for great demos.
The bet, the second time you read it, looks almost conservative. Software organizations already delegate. They delegate to junior engineers, to contractors, to other teams. Factory simply argues that some of that delegation will move to agents over the next decade, and that the company that gets enterprise trust first will own the pipe.
It is the same engineer, six months later. The Jira board has not gotten smaller - boards never do - but she has stopped dragging tickets one at a time. She has fanned out four Droids on four parallel branches. The Reliability Droid is already drafting the post-mortem for last night's pager. The Product Droid has turned three Slack threads into specs nobody asked it to write. Her coffee is still warm.
That scene is not science fiction. It is a Tuesday at one of Factory's enterprise customers. Whether the rest of the industry looks like it in 2028 is the open question - and the one Factory is now extremely well-funded to answer.
The droids you were looking for, it turns out, are already at work.