Factory.ai

Factory.ai, also known as Factory, builds AI systems called Droids to plan, build, review, test, and ship software. Founded in 2023 by Eno Reyes and Matan Grinberg, it focuses on enterprise software development, including migrations, maintenance, documentation, and incident response. Its platform expanded from specialized Droids into coordinated workflows with persistent execution, model routing, and analytics connecting agent sessions to delivered work.[2][12][10][9][11]

That broader platform followed earlier work on organization-wide automation and codebase understanding, beginning with the founders’ collaboration in 2023.[1][3][5]

Grinberg recalled meeting Reyes at a LangChain hackathon in 2023. Both had attended Princeton but had never had a one-on-one conversation; their encounter quickly turned to code generation, and they spent the next 72 hours building initial versions of Factory. Reyes was then at Hugging Face, advising customers’ technology and AI leaders and exploring systems that wrote and executed code. Grinberg was pursuing a theoretical-physics PhD at Berkeley and had shifted into AI research. He described code’s testability—the ability to verify an agent’s output against execution—as a central attraction.[1]

On November 2, 2023, Factory announced a $5 million seed round led by Sequoia and Lux, with participation from SV Angel, BoxGroup, Ali Ghodsi, Clem Delangue, Gokul Rajaram, and others. Its stated mission was “to bring autonomy to software engineering.” Droids were intended to automate organization-wide processes: planning features, fixing bugs, writing tests, reviewing code, and producing documentation.[3]

Factory announced a $15 million Series A on June 18, 2024, led by Sequoia Capital with Lux Capital and Mantis VC participating. It reported total funding above $20 million and a $120 million valuation. At that point, Factory described six specialist Droids—Review, Test, Code, Knowledge, Project, and Document—as operating at production scale, and named Bubble, Nav, Podium, and Stord among its customers.[4]

Code Droid executes coding tasks from natural-language instructions. Its June 2024 technical report describes HyperCode, a representation of an engineering system at multiple levels of detail. It combines explicit graph relationships with implicit relationships based on latent-space similarity, then extracts higher-level insights. ByteRank retrieves task-relevant information from those insights and underlying data, giving the Droid codebase context before it begins editing.[5]

Code Droid decomposes high-level problems into subtasks and uses development tools—including version control, linters, static analyzers, and debuggers—to obtain feedback while working. It can use different language models for different subtasks, generate multiple candidate trajectories, validate them with existing and self-generated tests, and select a solution from the candidates.[5]

Factory reported 19.27% on SWE-bench Full and 31.67% first-attempt success on SWE-bench Lite in June 2024. On Lite’s 300 problems, success rose to 37.67% at pass@2 and 42.67% at pass@6. During evaluation, Code Droid had no access to hints, internet results, test patches, or oracle patches, and HyperCode was frozen to the appropriate repository commit. Factory nevertheless acknowledged that some changes might have benefited from the underlying models’ exposure to the repositories during training.[5]

Factory also developed Crucible, a proprietary evaluation framework covering migration, refactoring, API integration, test generation, code review, documentation, and debugging. By May 2025, the founders said they no longer competed on SWE-bench because its tasks differed from their enterprise use cases. Reyes described internal code-editing evaluations built on Aider’s benchmark and behavioral evaluations with tasks, grades, and rubrics—for example, testing whether an agent asks questions when instructions are ambiguous. The 2024 report acknowledged the SWE-Bench and SWE-Agent teams and Albert Örwall’s SWE-bench-docker as sources of inspiration.[5][1]

By May 29, 2025, Factory’s product had reached general availability. In the founders’ interview, Reyes described the original Droids as asynchronous, event-triggered, relatively deterministic workflows. As models and Factory’s systems improved, Droids gained more freedom to choose their paths, guided by planning, decision-making, and environmental grounding. The demonstrated platform offered knowledge and technical-writing work, code execution, and reliability tasks such as compiling incident evidence and producing root-cause analyses.[2][1]

Factory’s enterprise approach gives agents context beyond source code. The May 2025 platform integrated tools such as Slack, Notion, Linear, Jira, and GitHub; it also generated codebase insights about project structure, development setup, and relationships between modules. Droids could propose a plan and ask clarifying questions before proceeding. Factory Bridge connected a Droid to a local development environment, with work able to move from local collaboration to remote execution.[1]

The founders initially chose a browser-based interface around delegation: developers could launch parallel tasks, inspect activity and context, and review completed changes. Grinberg argued that a workflow with less human code-writing would require more planning and verification. In May 2025, pricing combined small fixed access fees with usage billed through standard tokens, while enterprise reporting connected usage to pull requests, merged code, and code churn.[1]

A migration illustrates that delegation model. Reyes described analyzing a codebase, generating documentation and a migration plan, creating tickets with dependencies, then running independent tasks in parallel while a human reviewed changes and CI results. Grinberg reported that an unnamed large public company reduced a migration timeline from four months to about three and a half days. Factory worked alongside two customer engineers to demonstrate the workflow; those engineers then taught colleagues.[1]

Grinberg credited his older brother Cal with Factory’s early logo and work across brand, marketing, and product design. Reyes also credited the product-engineering team and described Droids using design systems to reproduce Cal’s brand voice and style.[1]

On September 25, 2025, Factory announced a $50 million Series B at a $300 million valuation, naming NEA, Sequoia Capital, J.P. Morgan, Nvidia, Abstract Ventures, and Mantis Ventures among the round’s leaders. Its product had expanded to CLI, IDE, Slack, Linear, and browser interfaces, with local or remote execution and model choice. Headless mode supported scripts and triggers customized to team workflows. Factory named MongoDB, EY, Bayer, Zapier, and Clari among organizations with enterprise rollouts.[6]

On April 16, 2026, Factory announced a $150 million Series C led by Khosla Ventures at a $1.5 billion valuation, with participation from Sequoia Capital, Blackstone, Insight Partners, Evantic Capital, 20VC, NEA, and Mantis VC. By then, it had introduced Missions for long-horizon, multi-step, multi-agent work and shipped Factory Desktop, giving Droids local context and system access. The founders reported daily use by hundreds of thousands of developers and revenue doubling month over month in each of the preceding six months.[7]

Factory acquired Lumetric on May 15, 2026. The YC W24 team, led by Ross Cefalu and Davis Palmie, had built model-agnostic AI workflows for deal teams. Factory said the team would help shape the next generation of its Desktop app, extending its investment in systems that understand context and work across tools.[8]

On June 1, 2026, Factory introduced Factory Router in private research preview for CLI and Desktop, including use by Mission workers. Router selects a model for each Droid session, escalates to a more capable model when necessary, and switches provider paths when one is unavailable. Administrators can supply routing guidance and govern model availability through organizational policies.[9]

Factory reported that Router retained 99% of Claude Opus 4.7’s pass rate at 20% lower full-session cost on Terminal-Bench 2, and 96% at 25% lower cost on Legacy-Bench. These were relative pass rates, measured across multiple runs using all 89 Terminal-Bench 2 tasks and the full Legacy-Bench suite.[9]

The founders’ June 15, 2026 Factory 2.0 article described a continuous software-development loop. Bug reports, internal conversations, customer feedback, and business requirements become planned changes; those changes are built, tested, reviewed, secured, shipped, and monitored. Monitoring produces new signals for the next cycle. The founders emphasized that organizational autonomy develops gradually through engineering work to codify workflows and standardize processes.[10]

Factory distinguished several levels of execution: simple Droids or skills for well-defined tasks; Automations for recurring workflows with shared objectives and memory; Droid Computers for persistent execution; and Missions that decompose complex work into parallel tracks over hours or days. The Desktop app added visibility for managing these processes. The platform described deployment options including hosted cloud, bring-your-own-key, a self-hosted data plane, EU-specific deployment, and air-gapped operation.[10]

Factory introduced Agent Effectiveness on August 13, 2026, in Private Preview. Its Throughput view compares cycle-time changes alongside Factory usage; Output maps spend to work intent; and Attribution links sessions to projects, issues, users, and artifacts. It requires connected work tools and the Advanced Analytics enterprise control. Output and Attribution cover only connected systems, and the cycle-time comparisons show associations rather than establishing causal productivity gains.[11]

On September 15, 2026, Factory announced $200 million raised at a $5 billion valuation, bringing reported total funding above $400 million. Financing participants included Blackstone, Khosla Ventures, Sequoia Capital, Insight Partners, Evantic Capital, Sound Ventures, NEA, Mantis VC, and Clearlake. Factory reported use by hundreds of thousands of developers and named Nvidia, Blackstone, Royal Bank of Canada, Palo Alto Networks, Adobe, and T-Mobile as organizations building software factories with its platform.[12][13]