# SWFT > Source-linked analysis and field guides on how product teams turn coding agents into dependable software, with bounded company cases and practical explanations. Publication updated: 2026-09-02 An AI software factory is a repeatable way to turn a clear product request into working software. People choose the goal and quality bar. Coding agents do bounded work inside a prepared system, and automated checks plus human judgment decide what can ship. SWFT publishes source-bound, AI-drafted guides and company case studies under the SWFT Editorial organization byline. Ben Guo's authenticated personal essay retains his Person byline. HTML is canonical. Every canonical page also has a text/markdown sibling, and the same representation is returned when the request explicitly prefers text/markdown. ## Start with the field analysis [Three visions of the AI software factory](https://swft.io/ideas/three-visions-of-the-ai-software-factory) separates three composable design layers that are often collapsed into one phrase: a bounded production line, an embedded outer loop, and a multi-agent work graph. It explains the accepted AI engineering terms, what each layer adds, and why most teams should begin with one measurable line. Browse [Ideas](https://swft.io/ideas) for arguments shaping the field, [Companies](https://swft.io/companies) for source-backed operating accounts, and [Guides](https://swft.io/guides) for practical explanations and references. ## Core pages - [The rise of software factories](https://swft.io/the-rise-of-software-factories): Ben Guo's author-owned personal manifesto. - [Three visions of the AI software factory](https://swft.io/ideas/three-visions-of-the-ai-software-factory): bounded production lines, Huntley's Ralph loop, and Yegge's multi-agent work graph. - [What is a software factory?](https://swft.io/what-is-a-software-factory): the plain-English working definition. - [How an AI software factory works](https://swft.io/software-factory-architecture): task specification, context engineering, agent harness, proof, release, and learning. - [How to build a software factory](https://swft.io/how-to-build-a-software-factory): a bounded pilot for product and engineering teams. - [Metrics](https://swft.io/software-factory-metrics): delivery time, rework, defects, human attention, cost, and product outcomes. - [Company cases](https://swft.io/companies): source-backed operating accounts with explicit limitations. - [Field map](https://swft.io/map): sourced operating cases, experiments, lineage, and platforms. - [Methods and corrections](https://swft.io/methods): evidence labels, source selection, and correction practice. ## Company cases - [Airbnb's LLM migration factory: 3,500 test files with a proof loop](https://swft.io/companies/airbnb-llm-migration): Airbnb split a 3,500-file test migration into independent work units with explicit stages. Models attempted each conversion, established tools checked it, and failures guided another attempt. The pipeline reached 97 percent before engineers used the partial work to complete the difficult final 3 percent. - [Amazon Q migrations: an AI factory for fleet-wide Java upgrades](https://swft.io/companies/amazon-q-migrations): Amazon Q Developer turned Java modernization into a repeated line: analyze an application, create a plan, change code and dependencies, build and test it, then ask a developer to review the result. Amazon reports fleet-scale savings, while the public evidence omits the full complexity mix and the product is moving toward end of support. - [Anthropic's AI-native SDLC: agents write, review, test, and investigate code](https://swft.io/companies/anthropic-ai-native-sdlc): Anthropic says agents author most of its merged code and participate across review, testing, and incident investigation. The reusable pattern is a controlled delivery loop around those agents: isolated computers, narrow identities, several kinds of checks, risk-based human approval, and records of consequential actions. - [Block's AI software factory: Goose, Builderbot, and the Buzz workspace](https://swft.io/companies/block-builderbot-goose-buzz): Block's factory has three layers: Goose is the open agent engine, Builderbot is the mature shared workflow from Slack or tickets to reviewed pull requests, and Buzz is an early multiplayer workspace for people and agents. The layers show how individual tools can grow into company infrastructure. - [Cloudflare's AI engineering stack: a control plane for company-wide agents](https://swft.io/companies/cloudflare-ai-engineering-stack): Cloudflare gives employees one controlled route to coding agents, models, company context, and internal tools. Identity, permissions, usage, cost, and review stay visible in shared infrastructure. Its adoption numbers are large, while the link from more activity to better product outcomes remains unproven. - [How Cognition uses Devin to build Devin](https://swft.io/companies/cognition-builds-devin): Cognition uses Devin across intake, code search, implementation, review, bug investigation, and recurring maintenance. Tickets, labels, schedules, and alerts can start bounded work. Reusable playbooks record outcomes and limits. Cognition reports high pull-request volume, but not the independent quality or productivity data needed for a causal claim. - [Cursor's agent factory: cloud computers, Bugbot review, and security loops](https://swft.io/companies/cursor-agent-factory): Cursor gives each cloud coding agent an isolated computer, asks it to test the running product, and returns visual and technical evidence with the code. Separate review and security agents check pull requests and recurring risks. Cursor reports substantial internal use, but the scale and outcome figures are not independently audited. - [Dropbox Nova: agents propose changes while deterministic systems control proof](https://swft.io/companies/dropbox-nova): Dropbox gives several coding agents one internal cloud environment. Each session starts from an exact code version, receives explicit checks, and can retry with real failure evidence. The agent edits and investigates, while deterministic systems keep control of continuous integration, branches, publication, and stop rules. - [How GitHub uses Copilot coding agent to build github.com](https://swft.io/companies/github-copilot-agent): GitHub uses Copilot coding agent in the private repository behind github.com. Engineers assign issues and receive pull requests that they can merge, revise, or close under the repository's existing controls. GitHub documents a broad task range and a lower merge rate than humans, but does not publish enough data to measure total productivity or quality. - [Linear coding sessions: from product issue to reviewed pull request](https://swft.io/companies/linear-coding-sessions): Linear coding sessions let a team delegate an issue where its product context already lives. Claude Code or Codex works in a managed sandbox, returns a diff and pull request, and can include verification artifacts. Linear also uses automated sessions on incoming bugs, although its outcome figure is a recent company self-report without a full quality or cost series. - [LinkedIn CAPT: turning company knowledge into executable agent playbooks](https://swft.io/companies/linkedin-capt): LinkedIn's CAPT layer gives compatible agents safe access to internal systems and step-by-step company workflows. More than 1,000 engineers use it, according to LinkedIn. Autopilot for Torch adds a specialist proof loop that keeps generating and checking machine-learning code until it meets explicit production measures. - [Meta's capacity agents: production signals become review-ready fixes](https://swft.io/companies/meta-capacity-agents): Meta starts agent work from measured production regressions or reviewed optimization opportunities. Shared tools gather evidence, reusable skills apply expert methods, and the agent prepares a fix for the responsible engineer. The company reports much faster investigations, while the public evidence does not isolate quality or total operating impact caused by agents. - [OpenAI's Codex factory: harness engineering and Symphony orchestration](https://swft.io/companies/openai-codex-symphony): One OpenAI team built an internal product by changing the repository and feedback systems around Codex, then created Symphony to assign project-board issues to isolated agent sessions. The public spec and operating account are unusually concrete, while the speed, scale, and adoption figures remain first-party evidence from bounded settings. - [Ramp Inspect: how a background coding agent became factory infrastructure](https://swft.io/companies/ramp-inspect): Ramp Inspect is an internal background coding agent that works inside a prepared remote development computer. Its important lesson is that the environment, company context, tests, attribution, and human review make the agent useful. Ramp published its architecture, while Inspect itself remains private. - [Replit's self-driving company: manager agents launch parallel work](https://swft.io/companies/replit-self-driving-company): Replit says each employee can use a manager agent to start several agents in parallel. Its wider system gathers context, checks results, and escalates work that needs human judgment. Agents run in isolated computers and reach approved company tools through controlled connections. The operating results are company self-reports, not an independent audit. - [Shopify River and Aquifer: one durable platform for many software agents](https://swft.io/companies/shopify-river-aquifer): Shopify reported that it was rolling Aquifer out profile by profile as one shared foundation for several kinds of agents. River works with employees in Slack. Aquifer is designed to preserve the session and supply temporary sandboxes, tools, credentials, and activity records. Roast adds repeatable checked workflows, and Dispatch shows how a specialist factory can require proof before review. - [Spotify Honk: adding an agent to a software factory that already worked](https://swft.io/companies/spotify-honk): Spotify already had rule-based systems for finding code repositories, applying repeated updates, opening review requests, and automatically merging well-understood maintenance work. Honk adds agent judgment where fixed scripts are not flexible enough, while the existing targeting, permissions, checks, and review stay in place. - [Stripe Minions: how developer infrastructure became an AI software factory](https://swft.io/companies/stripe-minions): Stripe Minions can take one delegation and return a review-ready code change because Stripe had already built the surrounding factory floor. Standard development computers, task-specific instructions, approved internal tools, fast tests, limited retries, and human acceptance make the unattended agent possible. - [Tempo's software factory: turning product signals into review-ready PRs](https://swft.io/companies/tempo-software-factory): Tempo routes approved product and error signals to Custom Agents that can work in parallel, open pull requests, and return results for human review. It is a direct small-company example of builders using their own product, with a useful decision loop and much less public technical evidence than the large-company cases. - [Inside Uber's AI software factory: measuring cost and quality at scale](https://swft.io/companies/uber-software-factory): Uber's public account shows how a large company can measure agent work as a system. It tracks accepted results, cost, quality, reusable skills, and managed workloads together. Google and METR add useful ways to evaluate behavior and autonomy, while every figure keeps its original evidence boundary. - [Vercel eve: the shared infrastructure behind more than 100 production agents](https://swft.io/companies/vercel-eve): Vercel eve is a shared framework for many kinds of production agents, including software work. It gives teams one file-based convention for resumable work, isolated computers, approved tool access, activity records, repeatable behavior checks, previews, human approval, and reversing a bad release. Vercel says more than 100 internal agents run on it. ## Structured surfaces - [Markdown sitemap](https://swft.io/sitemap.md) - [Atom feed](https://swft.io/feed.xml) - [JSON Feed](https://swft.io/feed.json) - [Software factory map JSON](https://swft.io/data/map) - [Glossary JSON](https://swft.io/data/glossary) ## Evidence labels - Observed artifact (OBS): an artifact or event readers can inspect. - Independent study (STUDY): research tied to a disclosed method and sample. - First-party report (SELF-REPORT): an organization or person describing its own work or results. - Analysis (INFERENCE): SWFT's synthesis of named evidence. Use the sources and limits on each page. Inclusion is not endorsement. Machine-readable surfaces contain the same admitted records as the visible publication, including visible reader questions when a page has them. There is no hidden answer-engine copy.