Factory.ai and Droid: what is a software factory and how agents go through the entire development cycle from ticket to release
Main chat
A chat for vibe coders: news, guides, live cases, marketplace, and finding executors.
Most AI development tools solve one problem: they write code faster than humans in the editor. Factory.ai was built around a different idea from the very beginning: what if the agent had to do more than just finish the function, but run the task through the entire life cycle, from the problem signal to the release, monitoring and documentation, so that the system gets smarter with each cycle.
Founded in 2023 by Matan Greenberg (a graduate of the PhD program at Berkeley) and Eno Reyes, the company built a platform around autonomous agents named Droid – and in April 2026 closed the Series C round at a $150 million valuation of $1.5 billion, ranking alongside Cursor, Cognition (Devin) and Claude Code from Anthropic as one of the most capitalized players in the autonomous development market.
What is Factory: From Tool to Software Factory
The company formulates its position through the metaphor of production: the goal is to build a self-improving system for the entire life cycle of software development (SDLC), which absorbs a continuous stream of signals and releases ready-to-production software.
The idea differs from the usual “AI assistant in the editor” fundamentally. According to the description of the company, the software plant begins with signals from the outside world - bagreports, internal discussions, customer reviews, business requirements. These signals go through triage and turn into planned changes. Next, the task goes through stages: signal → triage → plan → execution → verification → release → monitoring → automation – and each stage strengthens the others, because they all work on a common agent core with a common organizational context.
The company’s co-founder explains the philosophy that most AI development tools fail in production not because the models lack capabilities, but because the corporate context is too chaotic for an agent to work reliably. Therefore, Factory focuses on what it calls “road laying” – integrating Droid with company documentation, test coverage infrastructure, CI/CD pipelines, internal APIs, and compliance tools before an agent writes a single line of code.
Droid: What an agent is built on
Independence from the model
Factory’s key architectural solution is to refuse to be attached to one model. The platform supports all leading advanced and open models – including current versions of the GPT-5, Claude Opus and Sonnet, Gemini and others.
This is not just a marketing phrase about choosing a model in settings. One Droid task can pass the baton between Anthropic models for scheduling, DeepSeek for volumetric code generation, and a small open model for routine test writing in a single workflow. Factory treats model selection as a routing task rather than a brand issue.
Where does Droid work
Droid is available in several surfaces with common context and common settings: a command line interface, editor extensions (VS Code, JetBrains, Vim), a web dashboard for viewing difs and task history, a desktop application for macOS and Windows, a mobile version for viewing progress and approving changes, and cloud-based “computers” – managed virtual machines on which Droid can work autonomously without straining the local developer environment.
A telling detail from the company’s own telemetry: Teams using both the command-line interface and the desktop app are deploying Droid twice as fast, and those working in both environments are performing 4.6 times as many sessions as command-line users — meaning people aren’t splitting time between interfaces, but doing more work together.
Results from independent testing
In the Terminal-Bench open industry test, which measures the ability of agents to perform complex tasks in a command-line environment, Droid, with a score of 58.8%, ranks first among all agents tested. It is significant that the architecture of the agent, and not only the choice of the model, turns out to be the decisive factor: with the same basic model, Droid outperforms its own agent tools of laboratories - for example, Droid on the Opus model shows 58.8% versus 43.2% for Claude Code on the same model, and Droid on the GPT-5 - 52.5% versus 42.8% for Codex CLI.
It is worth noting the honest detail that the company itself admits: Factory previously published the results on the SWE-Bench benchmark, but at some point stopped doing this, explaining that the test is limited only to Python and does not correlate well with real corporate tasks – migration of legacy systems, writing code in compliance with audit requirements and orchestrating multi-agent workflows. It’s a reasonable argument, but it’s also convenient for a company that doesn’t have fresh public results on this particular test — it’s worth keeping both facts in mind at the same time.
Missions: the main feature of the platform
One thing that sets Factory apart from similar tools is Missions: structured multi-agent workflows for large projects that span hours to days.
How it works architecturally
Inside the mission are agents with separate roles. **The organizer plans and breaks down the approach to achieving the user's goal, directs the execution to completion and passes all verification barriers - but does not engage in small routine work, delegating the investigation and implementation to the workmen. Workers perform well-defined functions with clear criteria for success, iterate not yet sure of the correctness of the result, and transfer the work further - but the final decision on the correctness is not theirs. **Validators is an independent authority that evaluates the completed work for correctness and completeness, identifies bugs and gaps, but does not itself correct: instead, transfers the problems found back to the orchestrator, who creates correction tasks for future workmen.
After each check, the orchestrator revises what the workmen and validators have noted, creates targets to correct the identified problems, and the cycle is repeated until the next stage (milestone) is passed. If the execution or verification comes to a dead end, the mission stops and the control is returned to the person.
Real statistics on how missions work
Figures from the company’s proprietary telemetry give a visual sense of scale: a typical interactive session with Droid lasts an average of about 8 minutes, and 60% are completed in 15 minutes. Missions are arranged fundamentally differently: the median duration is about 2 hours, 65% last longer than an hour, 37% longer than four hours, and 14% last longer than a day. The longest mission recorded lasted 16 days of continuous autonomous operation.
At the same time, missions spend resources in a completely different way than normal sessions: in the median, the mission consumes 12 times more tokens than a normal session, and at the 99th percentile, the gap is 9 times – although the rate of token consumption per minute remains about the same (about 45 thousand tokens per minute), just the mission keeps this pace much longer.
Specific example of the result
The company gives a good example: one mission independently created a clone of Slack – authorization for workspace, channels and discussions, real-time messaging with reactions and mentions, file downloads, search, presence statuses and notifications. The project passed six control stages, the verification took 37.2% of the total execution time, the final result is 38.8 thousand lines of code, more than half of which are tests, with coverage of code operators 89.25%.
Launch of the mission
To start a mission, the command /missions or /enter-mission is enough in any Droid session – after that, the agent asks clarifying questions, probes the limitations and iterates the plan with the user before proceeding to execution. Headless launch without interactive dialogue is available via droid exec --mission.
Sovereign Deployment: Control to Isolated Networks
For corporate and government customers, one of the crucial factors in choosing a tool is where the data and code are physically located. Factory builds around this a whole range of deployment options: a fully managed cloud without the need for its own infrastructure, a hybrid version with a control plane in the Factory cloud while maintaining client-side computing, deployment entirely in the client’s own data center, and a completely isolated environment without any Internet access – with support for major cloud providers (AWS, Azure, GCP) where the client needs it.
Sovereignty is not just about choosing where a system physically works. It is the possession of a system that learns on its own, feeding every agent session, every code review, and every incident resolved back into the overall learning cycle – and yet this accumulated ability remains within the client’s walls, under their control.
Air-gapped deployments are already operating at some of the largest financial, medical and government organizations in the world for which code cannot physically leave the perimeter of the organization.
It also supports the option of fully local models: you can register a machine with a GPU, connect through Ollama, vLLM or any compatible endpoint – and Droid will work entirely on local models without transferring any data outside the internal network.
Context and Memory: How an Agent Differs from a Chatbot with Repository Access
A separate architectural layer is organizational and user memory. The platform captures decisions, documentation and accumulated work practices so that each agent remembers the context between sessions - without having to re-clone and analyze the entire repository from scratch each time.
Native integrations with GitHub/GitLab, Jira, Slack, PagerDuty, and real-time indexing mean that each agent sees the same architectural diagrams, tickets, and code that a human developer sees on the team – that’s the “context before code” the company talks about.
Prices: four levels and billing token
Factory’s tariff structure is built around resource consumption, not the number of jobs — an important difference from the user-to-user peg typical of most SaaS products.
Pro - $20 per month, minimum entry plan for the product. Factory does not have a free rate for trial evaluation – this distinguishes it from Cursor (there is a free level) and Claude Code, which is included in the Claude Pro subscription.
Plus – $100 per month, a significant jump in available resources relative to Pro.
*Max $200 per month, roughly ten times the use of the Pro plus early access to new features, including Missions.
Teams is an individually agreed price for teams up to 150 jobs, with single authorization (SSO), SAML/SCIM management, zero data retention option and basic administration tools.
Enterprise - individual price, unlimited number of team members, dedicated computing resources with an isolated inferencing pool, logging actions for audit, deploying on its own infrastructure and a full set of administration tools.
Billing is arranged through spendable loans (Standard Usage), with three independently rotating limits - for 5 hours, 7 days and 30 days. When the standard limit is exhausted, you can continue working on the models of the Droid Core line - a set of open models designated by the company that effectively cope with most programming tasks at a low cost, and use a separate pool of limits. After both pools are exhausted, a prepaid use payment (Extra Usage) is available with a minimum payment of $10.
An important detail for missions is that they consume the same rotating quota as regular sessions, but with the use of multipliers for the selected model, and to start them, the Extra Usage option must be included - if the mission rests on the limit in the middle of the path, it simply pauses, and not ends with an error.
Who uses Factory: the scale of the customer base
According to the company, Droid is used by hundreds of thousands of developers in organizations such as Nvidia, Adobe, Bayer, EY (Ernst & Young), MongoDB, Zapier, Adyen, Palo Alto Networks, Framer, Groq, Chainguard, Podium and Clari.
The Series C round in April 2026 was led by Khosla Ventures, with participation from Sequoia Capital, Insight Partners, Blackstone and several smaller funds. The company’s revenue has reportedly doubled month-to-month for six consecutive months, with hundreds of thousands of active daily developers now working through large enterprise customers.
One fintech client, cited by the company itself, describes the motivation behind choosing Factory precisely through a data security prism: the company was concerned about how to use AI while maintaining strict privacy and security standards, and Factory gave a controlled way to unify the engineering context without compromising on compliance requirements - which is fundamentally different from a situation where sensitive information leaks into arbitrary language models without any control.
Factory is different from Cursor, Claude Code and Devin
It is useful to fix the positioning difference rather than treating all of these tools as interchangeable variants of the same thing.
Cursor, Windsurf, GitHub Copilot – focused on an individual developer and small teams: fast auto kit and chat assistant right inside the editor, focus on response speed at the time of writing code.
Factory, Devin by Cognition, Claude Code for Enterprises are focused on standalone tasks with a long horizon that span multiple repositories, systems and real-time days without a constant human presence at every step.
The distinction is important to the buyer: a pay-as-you-go tool for individual development and a stand-alone agent are not interchangeable categories, and large enterprises are now increasingly buying both at the same time rather than choosing one over the other.
Specifically from Claude Code and Codex, the CLI Factory is distinguished by its explicit emphasis on multiple models within a single task (rather than being tied to a single model family), on structured multi-day missions with orchestrator/worker/validator roles segregated, and on a sovereign deployment infrastructure designed for regulated industries from the start rather than as an option added later.
The honest limitations of the platform
Independent review materials, not related to the company itself, note several practical limitations that should be considered before implementation.
**There is no free rate for trial evaluation. ** Unlike Cursor with its free tier or Claude Code, which is available as part of a Claude Pro subscription, the minimum fee for Factory is $20 per month without the option to test the product for free.
** Token billing requires discipline. ** Unlike fixed pay for jobs, token billing transfers the responsibility for controlling costs to the user himself - for disciplined teams this provides a profitable cost, but requires constant attention to consumption, rather than a one-time decision to buy a subscription.
The product is not for those who want “just a faster auto kit”. ** The very nature of the tool is designed for structured teams that are ready to build code review and verification processes around the agent – if the team is not ready to maintain the discipline of viewing the results of the agent’s work, there is a risk of either stinking low-quality code or wasting tokens on useless alterations.
Droid is not immune to the typical problems of generative AI. Like any other language model tool, an agent can hallucinate logic, miss boundary cases, or write ineffective solutions—especially if the original prompt or repository itself is vaguely formulated.
Bottom line: for which team does it make sense
Factory.ai is not a tool for a solo developer who wants a faster auto kit, nor is it a replacement for Cursor for everyday code editing. It is a platform designed for teams ready to build a full-fledged process around the autonomous agent: from the first signal of the problem to the release and subsequent monitoring, with transparent control at every meaningful step.
If your team already works with structured ticketing, has a customized CI/CD, values choosing the best model for a specific task instead of being tied to a single provider, and – especially important for regulated industries – needs a deployment option up to a completely isolated environment, Factory deserves serious consideration. If you want a fast and inexpensive assistant inside an editor without the complex infrastructure around it, this is probably not the task Factory does best.
*This article is based on official Factory.ai (factory.ai/news, docs.factory.ai), independent reviews of The AI Agent Index, Fritz AI, Developers Digest, and funding round materials from Tech Insider and Idlen. Actual on June 23, 2026. *
