For three years, working with a coding agent has meant writing tickets for it. Cognition, the $48 billion company behind Devin, now wants engineers to stop doing that. On November 17 it launched Devin Bots, persistent agents with their own memory and tools that take on an ongoing goal, such as “keep Android at parity with iOS,” and then lead teams of Devin sessions to deliver it. Cognition calls it its biggest launch of 2026, and its showcase is bold: an Android app built in 44 days, with one engineer sending the Bot 183 messages. Here is how Devin Bots work, what the numbers do and do not show, and where the idea fits in the crowded race for always-on agents.
Quick Take
- What: Cognition launched Devin Bots, persistent “engineering teammates” that own a goal, plan the work, start parallel Devin sessions and report back.
- How: Each Bot has one chat, its own memory, connections and secrets, can run on schedules or triggers, and collects decisions for humans in a “Needs You” inbox.
- Showcase: Cognition says an “Android Bot” built Devin’s Android app in 44 days, merging 1,134 pull requests and writing 344,651 lines of Kotlin.
- Where: Bots work in the Devin app, Slack (with their own identity) and iMessage, and can coordinate with each other.
- The catch: All headline results come from Cognition’s own use, and humans still define success, set permissions and approve sensitive actions.
In plain English
Until now, AI coding assistants mostly worked like contractors who need a precise instruction for every job. Devin Bots are meant to act more like a team lead: you give one a goal, and it decides which tasks are needed, hands them to other AI helpers and checks the results. It remembers what happened before and only asks a person when a real decision is needed. Cognition says this let one engineer get an entire Android app built with a few messages a day, but these results come from the company itself.
What happened
Cognition announced the product on X and in a blog post, “Introducing Devin Bots: from tasks to outcomes”:
Meet Devin Bots.
— Cognition (@cognition) November 17, 2026 · View on X
From writing firmware to merging thousands of PRs a week, a Devin Bot leads a team of agents to get the job done.
“A Bot is a persistent agent with its own memory and tools that can lead multiple Devins in parallel,” Cognition writes. “It retains context across sessions and brings you only the decisions that need your judgment.” Devin Bots are available now via the Devin website.
Jeff Wang, Cognition’s President of New Enterprise and former Windsurf CEO, framed the launch as the next step after autocomplete, local agents and cloud agents: “Today is our biggest launch of 2026.” The launch follows Cognition’s $2 billion Series E at a $48 billion valuation in September.
Background: from tasks to outcomes
Cognition’s core argument is that a coding agent “can finish every task you give it while still leaving you responsible for moving the project forward.” Someone still has to spot the next problem, choose an approach, write the task and check the result. Devin Bots try to take over that coordination layer. The blog contrasts the two styles: “Port this feature to Android” is a task; “Keep Android at parity with iOS” is a responsibility.
The idea is not unique. Persistent, always-on agents are a crowded field, with OpenAI’s dots, Anthropic’s Duet and Grok Bot all pitched as teammates rather than tools. Grok Bot has already been used this way, for example by the Grok Bot team that shipped 11,000 pull requests in a month. Cognition’s angle is that Devin Bots are built specifically for software engineering, with Devin’s existing integrations and permissions. Cognition also acquired The Interaction Company, maker of the personal assistant Poke, in July; Bots engineer Samyok described Devin Bots as a continuation of “the ‘human-like’ feel from Poke, but for engineering.”
How Devin Bots work
According to Cognition’s documentation:
- Chat: Each Bot has one chat for everything it owns. You describe the outcome there and its reports come back there.
- Sessions: Each piece of work runs as a child Devin session with its own cloud machine. The Bot starts and monitors these sessions.
- Automations: A Bot can run on a schedule or when something happens in connected tools, without a message from a person.
- Needs You: Approvals, decisions and pull requests to review are collected in one list.
- Memory: What the Bot learns is saved and used in later sessions.
- Access: A Bot has its own connections and secrets, separate from the people who message it, and owners choose who can use it.
Cognition stresses that engineers still define success criteria, build the feedback systems (tests, benchmarks, production metrics) and decide “which systems it can access, what it can change, and when spending or releases require approval.”
The three showcases
- Android app: On September 28, engineer Samyok told a new Android Bot: “your job is to product manage the android app.” Over 44 days, Cognition says, the Bot merged 1,134 pull requests, wrote 344,651 lines of Kotlin, added 2,841 screenshot tests and closed 1,217 tickets. Samyok sent 183 messages, “about one for every six pull requests it merged,” and none on 22 of the days. The app is in open testing on Google Play.
- Layout Bot: A Bot that runs a layout check every three hours on the latest code, traces regressions to the pull request that caused them and opens fixes with before-and-after videos. Cognition says it merged 159 pull requests from the Bot in three weeks, “cutting lag by 70%.” Asked who had requested one fix, the Bot replied: “Nobody asked.”
- Game studio: Cognition employee Jenny Xu runs a mobile game studio with about 20 Bots and one “Super Jenny Bot” that collects decisions for her four times a day. Cognition says the studio went from about 50 to 1,900 pull requests a week.
Pros and cons
For users (engineers)
- Pro: Less time writing tickets and chasing agent sessions; decisions arrive in one inbox, and Bots can be reached via Slack or iMessage.
- Con: Reviewing a constant stream of Bot-generated pull requests becomes the new bottleneck, and trusting auto-merged code requires strong tests.
For developers and tool builders
- Pro: A clear model for agent management (owner, scope, memory, permissions) that others can learn from.
- Con: The workflow is tied to Devin’s cloud and pricing; it is not an open standard.
For companies
- Pro: Neglected but valuable work (performance, parity, maintenance) can get a permanent owner without new headcount.
- Con: Persistent agents with their own secrets and broad repository access widen the security and governance surface.
What X is saying
Most launch-day posts came from Cognition staff, investors and beta testers, so the tone was overwhelmingly positive. CEO Scott Wu put the pitch as a math problem:
Math time: if a coding agent goes from doing 98% of your work to doing 99%, how much more can you get done?
— Scott Wu (@ScottWu46) November 17, 2026 · View on X
It can be hard to explain a new product experience in words. But the difference we’ve seen both with beta testers and internally @ Cog is enormous. Give it a try and let us know what you think 🙂
Developer and podcaster swyx (@swyx), who said he had been trialing Bots for a week, called them “multi-agent teams done right,” adding that his “main aha moment is that you don’t want ‘AI Employees’” but “agent managers that own outcomes.” Cognition’s Nader Dabit (@dabit3) wrote that he had “migrated 100% of my work to Devin Bots over the past few weeks.” Brand lead Joseph Alessio noted that the Bots’ visual identity was partly generated by Bots themselves. Other users pointed out how crowded the field has become, listing Grok Bot, OpenAI’s dots and Anthropic’s Duet alongside Devin Bots. We did not find a Grok summary or reply specifically about Devin Bots at the time of writing.
Critical analysis
Devin Bots address a real gap. As agents get better at single tasks, the human cost shifts to coordination: deciding what comes next, checking results and remembering context. Packaging that as a persistent, permissioned “owner” is a sensible design, and the Android example, if representative, is genuinely striking.
But the evidence base is narrow. Every number comes from Cognition’s own teams or its employees’ projects, under ideal conditions: engineers who know the product, generous compute and a culture that already merges more than 2,000 pull requests a week. Pull request counts and lines of code measure output, not quality or maintainability; 344,651 lines of Kotlin in 44 days will also need maintaining. The launch-day chorus, much of it from staff and beta testers, adds little independent signal. The real test is whether ordinary engineering teams, with messier codebases and stricter review rules, see the same leverage without drowning in review work.
What’s still open
- Pricing and usage limits for running many persistent Bots and child sessions.
- Independent case studies from customers outside Cognition.
- How Bots handle conflicting goals, failed deployments or security incidents without human input.
- Long-term code quality and maintenance cost of Bot-built software such as the Android app.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent i10x analysis is based on Cognition’s announcement, product page and documentation, and public posts on X by Cognition, its staff and developers, as of November 18, 2026. All performance figures are as reported by Cognition.
Sources (plain text, no links):
- Cognition: Introducing Devin Bots: from tasks to outcomes (November 17, 2026)
- Devin documentation: Devin Bots
- Cognition: Devin Bots product page
- Posts on X by Cognition, Scott Wu, Jeff Wang, swyx, Nader Dabit, Samyok and Joseph Alessio (November 17, 2026)
i10x Perspective
The agent race has moved up a level. A year ago the question was whether an AI could close a ticket; now it is whether an AI can own a backlog. Devin Bots are one of the most concrete answers so far, and the “responsibility, not task” framing is likely to outlast this particular product. For teams considering it, the sensible start is a contained responsibility with clear metrics, such as a performance check or a parity project, and a hard look at how much review time the Bot’s output actually costs. More on the tools reshaping software work in our Coding section.



