--- schema: "swft.publication/v1" id: "company-cognition-devin" title: "How Cognition uses Devin to build Devin" description: "How Cognition uses Devin for code search, pull requests, automated review, design-system audits, bug triage, playbooks, and workflow learning." summary: "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." canonical: "https://swft.io/companies/cognition-builds-devin" author: "SWFT Editorial" author_type: "Organization" published: "2026-09-01" modified: "2026-09-02" kind: "case-study" section: "Companies" tags: ["Cognition Devin", "Devin Review", "coding agent playbooks", "automated bug triage", "AI code review", "AI software factory"] evidence_labels: ["INFERENCE", "SELF-REPORT"] source_ids: ["cognition-builds-devin"] authorship_disclosure: "AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections." --- # How Cognition uses Devin to build Devin Cognition starts Devin from chats, tickets, alerts, schedules, and APIs, then connects implementation to review, CI repair, bug investigation, and reusable playbooks. > **Authorship:** AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections. ## Quick answer 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. Cognition uses Devin to build Devin through several repeatable workflows. A request can begin in Slack, Linear, the web app, a command line, or an API. Devin searches the codebase, carries out the task, prepares a pull request, responds to checks, and returns the change for review. ## One agent, several ways to start work Cognition's [February 2026 operating account](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin) says a person can tag Devin in a chat or ticket, attach supporting material, and receive a pull request to review and test. Employees do not need a local repository or command-line setup to begin a bounded contribution. For less defined work, Ask Devin searches an indexed version of the codebase before a coding session begins. The person can clarify the goal through that exploration, then pass the gathered context into a session. This separates **scoping**, which determines what should change, from **implementation**, which changes it. Cognition reported 659 Devin pull requests merged during the week before publication, compared with 154 in its best week of 2025. That is a large first-party volume claim. The article does not report the size of those pull requests, the share of all merged work, the number of failed attempts, review time, defects, or customer outcomes. ## Review is a separate production line Cognition says Devin Review runs on every internal pull request. It reorganizes large diffs into related groups, explains changes, detects moved code, flags possible bugs by confidence, and lets reviewers ask codebase-aware questions. When Devin Review or another GitHub check finds a problem, Autofix can update the pull request. Devin also works through CI and lint failures until the required checks pass. CI, or continuous integration, is the automated build-and-test process that checks a proposed change before it merges. This loop reduces handoffs between generation and repair. It does not remove the human checkpoint: Cognition's account says employees review and test the pull request. A passed check shows that a defined condition holds. It does not prove that the change solves the right customer problem or has no defect outside the tested conditions. ## Schedules and events turn policy into recurring work Cognition runs a daily design-system audit. Every morning, Devin checks pull requests merged during the prior 24 hours for hardcoded colors, nonstandard spacing, and components that should use the shared library. It creates Linear issues and can prepare fixes. The trigger makes the policy repeatable. The system does not depend on a designer noticing each violation or remembering to ask for a cleanup. The output is also bounded: a finding, an issue, and sometimes a proposed pull request. Bug triage begins when someone applies a Bug label in Linear. A playbook tells Devin to read the report, find relevant code, inspect Git history, and return a likely root cause and repair plan. With access to Datadog and a read-only database replica, a deeper debugging job can inspect logs and data, trace the breaking change, write a regression test, and prepare a fix. Read-only production access matters. The agent can gather evidence without receiving permission to change live customer data. A regression test records the failure as an executable condition, so the proposed fix and later changes can be checked against it. ## Playbooks make a successful session repeatable Cognition defines a Devin Playbook as reusable guidance for work that appears repeatedly. Its recommended fields include the outcome, required steps, postconditions, known model tendencies, forbidden actions, and context the requester must provide. A **postcondition** is a fact that must be true after the task finishes. It can be more useful than a long instruction sequence because it leaves the agent room to choose an approach while keeping the result checkable. Cognition reports using playbooks for data ingestion, database migrations, and integrations with services such as Stripe, Plaid, and Modal. Session Insights examines completed work for problems, scope changes, milestones, and possible prompt improvements. A later session can start with that revised guidance. This is a learning loop, although Cognition does not publish a controlled measure showing how much those recommendations improve later results. ## Public access stops before the internal operation Devin is a commercial product, and DeepWiki provides public codebase documentation plus a free MCP connection. Cognition's private repositories, prompts, playbooks, review data, success rates, and production connections are not public. The article comes from the company that develops and sells Devin. The 659-pull-request figure covers one week and provides no independent quality measure. It shows that Cognition uses Devin at meaningful volume. It does not establish that another company, task mix, or repository would achieve the same results. ## What a product team can copy Let work begin where the evidence already appears, such as a ticket, alert, or pull request. Separate scoping from implementation. Give production investigators read-only access where possible. Turn a successful recurring task into a playbook with an outcome, postconditions, forbidden actions, and required inputs. Keep review as its own workflow, and measure failed attempts, review time, defects, and accepted outcomes alongside pull-request count. ## How we know - **First-party report (SELF-REPORT)** Cognition says it uses Devin through web, Slack, Linear, command-line, and API entry points and merged 659 Devin pull requests in one reported week. Sources: [How Cognition Uses Devin to Build Devin](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin). - **First-party report (SELF-REPORT)** Cognition describes Devin Review on every internal pull request, automated responses to review and CI findings, and a daily design-system audit that creates issues or fixes. Sources: [How Cognition Uses Devin to Build Devin](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin). - **First-party report (SELF-REPORT)** Cognition describes event-triggered bug triage, read-only production investigation, reusable playbooks, and session analysis that can improve later prompts. Sources: [How Cognition Uses Devin to Build Devin](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin). - **Analysis (INFERENCE)** SWFT treats explicit triggers, read-only investigation, postcondition-based playbooks, and separate review loops as the most transferable parts of Cognition's system. Sources: [How Cognition Uses Devin to Build Devin](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin). ## Sources - **First-party report (SELF-REPORT)** [How Cognition Uses Devin to Build Devin](https://cognition.com/blog/how-cognition-uses-devin-to-build-devin) — Cognition; published 2026-02-27; accessed 2026-09-01. Cognition's detailed internal workflow from Slack, Linear, and event triggers through sessions, review, automated repair, recurring audits, and pull requests. ## Read next - [Stripe Minions: how developer infrastructure became an AI software factory](/companies/stripe-minions) - [Ramp Inspect: how a background coding agent became factory infrastructure](/companies/ramp-inspect) - [Software factory metrics that matter](/software-factory-metrics)