An agent that takes a goal instead of a task

Cognition, the U.S. company behind AI programmer "Devin," announced "Devin Bots" on Nov. 17 (Nov. 18 Japan time). The company describes a Devin Bot as an agent that is given a goal by the user and leads a team of Devins to deliver it. The official blog post called a Bot a persistent agent with its own memory and tools that can lead multiple Devins in parallel.

Meet Devin Bots. From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.

The announcement came in an official blog post dated Nov. 17 and a post on the company's official X account (2:03 a.m. Nov. 18 Japan time). According to the blog post, a Bot takes on a large project or an ongoing goal, investigates on its own, delegates work, and evaluates the results to decide what to pursue next. It keeps context across sessions and brings the user only the decisions that need a human, the company said.

Until now, the company has expanded Devin around the model of Devin handling individual tasks. In November it announced "Devin Review 2.0," which checks code, and "Devin Apps," which lets companies share web apps built by Devin internally, and it also revamped its "Teams" plan for groups. The blog post says coding agents have so far been handed specific tasks, while engineers remained responsible for coordination: noticing a problem, choosing an approach, checking the result and deciding what comes next. Cognition positioned Devin Bots as a product that takes on that coordination work.

The blog post illustrated the difference with three examples. Where a task would be to port a feature to Android, a Bot is given the responsibility of keeping Android at parity with iOS. In the same way, fixing a layout shift becomes keeping the web app smooth as it changes, and fixing an onboarding error becomes getting new users through setup successfully. Other examples of work a Bot can own include maintaining an open-source repository, managing releases and optimizing server inference.

In-house case built an Android app in 44 days

The blog post described three cases in which the company gave Bots responsibilities internally. All of the figures are Cognition's own, and none have been verified by third parties.

The first was building Devin's Android app. According to the company, Devin had an iOS app and iOS engineers behind it, but no Android engineers and no plan for Android. On Sept. 28, 2026, Samyok, an engineer at the company, created "Android Bot" and told it to product-manage the Android app and build the whole thing autonomously. The Bot first raised blockers, such as where the code would live, how to test an Android build and how changes would be merged, and Samyok set the direction and constraints. From there, the Bot filed tickets, started a Devin session for each one, and merged pull requests (proposed code changes) that passed CI (automated testing) and "Devin Review." It posted a daily report with a screen recording in Slack, the company said.

Forty-four days later, Devin had an Android app. According to the company, over that period the Bot merged 1,134 pull requests, wrote 344,651 lines of Kotlin, added 2,841 screenshot tests and closed 1,217 tickets. The app is in open testing on Google Play, and the Bot updates it every time a new feature lands in the iOS version.

Humans remained involved. Samyok sent the Bot 183 messages over the 44 days, about one for every six pull requests it merged. On 22 of the days, no messages were sent. Most of the messages responded to matters that needed a human decision, such as a pull request requiring a security review, the company said.

Also used against layout shifts and to run a game studio

The second case was "layout shift," in which elements of the Devin web app move on screen when content loads late. The company created "Layout Bot" and gave it the responsibility of finding and fixing these shifts. According to the blog post, every three hours the Bot builds the latest code on the main branch the same way as in production and records first visits, repeat visits and loads with individual requests deliberately delayed. When a previously passing check fails, it traces the regression to the pull request that introduced it, has a Devin session prepare a fix, and sends it to an engineer for review with before-and-after recordings.

Cognition said it merges more than 2,000 pull requests a week. According to the company, in the three weeks Layout Bot has been active, 159 of its pull requests were merged, and lag was cut by 70%. The company did not say which metric the 70% reduction refers to or what values it fell from and to.

The third case was a mobile game studio run by Jenny Xu, who is now a Cognition employee. According to the blog post, Xu is the studio's only human employee, and about 20 Bots each own one job, including development, Unity, releases, UI, iOS, Android and testing on real devices. Above them, a "Super Jenny Bot" collects the decisions the other Bots need into a single Slack thread four times a day, and Xu answers with emoji. The studio went from about 50 pull requests a week to 1,900 and ships a release every week, the company said.

The blog post also carries comments from companies that used Bots early. Bilt's vice president of engineering said, "Last week, our Devin Bots merged about 1,600 PRs." All of the comments were selected and published by the company.

Humans set the bar for success and the limits of authority

The blog post says that, as a starting point, engineers need to define what success means for the scope they give a Bot. For example, a Bot responsible for making CI faster needs to know which checks must remain and whether spending more on compute is acceptable. Engineers also build the way a Bot evaluates its work. While a product is being built, that means tests, benchmarks and running through user flows; once it is in use, it means production behavior, usage and customer feedback, the company said.

Engineers also decide which systems a Bot can access, what it can change, and when spending or releases require approval. Bots proactively share reports and test recordings, and decisions that need input land in an inbox called "Needs You." Users can open individual Devin sessions to dig deeper, and Bots can be shared with a team. When multiple Bots are in place, they coordinate with one another to bring their updates and decisions together, the company said.

Available the same day, no price given

The blog post said "Devin Bots are available today at devin.ai/bots.", launching the product alongside the announcement. Neither the blog post nor the product page gives a price, the plans that include Bots or usage limits.

According to the blog post, users can work with Bots in Slack, iMessage or the Devin app. The product page also lists text messages, calls and group chats. Bots can be set to run automatically on schedules and triggers, such as watching Slack for customer issues, checking CI health nightly and monitoring Linear for blockers. According to the product page, Bots use Devin's existing capabilities, including codebase understanding, a cloud computer for every agent, code review and automated testing.

The product page offers six ready-made templates. They are "Software Factory," which keeps opening small improvement pull requests; "Android Engineer," which builds Android features; "Chief of Staff," which coordinates the team; "Latency Optimizer," which speeds up slow APIs; "Product Engineer," which builds features and fixes bugs; and "QA Engineer," which tests pull requests in a real browser. For businesses, the page lists SOC 2, ISO 27001 and FedRAMP, along with zero data retention with model providers, audit logs, a dedicated virtual machine for each session and single sign-on, among other features.