AI becomes the boss of AI: Devin Bots run teams of AI coders for weeks

Startup Cognition has unveiled "Devin Bots": an always-on AI agent that directs teams of Devin coding agents over days or weeks. At one game studio, weekly code changes reportedly jumped from 50 to 1,900. Pricing and general availability haven't been announced.

Projects Nov 17, 2026 8 min read Source: dev.to #ai
Illustrative image, AI-generated

What's going on

Cognition, the US startup behind the AI coder Devin, unveiled Devin Bots on Monday, November 17, 2026. Put simply, the AI software engineer now gets a manager, and that manager is an AI too. A Devin Bot is an always-on agent, meaning an AI program that works through several steps on its own, one after another. It oversees entire teams of Devin agents on complex projects that run for days or weeks.

Cognition gives an example. You ask the bot to build an Android version of an existing iPhone app and keep it in sync whenever the iPhone version changes. The bot breaks the work into subtasks, hands them out to individual Devin instances and keeps track of the whole project.

For now, only a few dozen customers can try it, in early access (a closed testing phase). Fortune broke the news as an exclusive. There are no independent hands-on tests yet. The author of the main source on dev.to says plainly that he couldn't use the product himself.

From solo worker to project lead

The original Devin launched in March 2024 and was pitched as the "first AI software engineer." It was built to work like a single employee: you give it a ticket, meaning a specific task, and it writes the code. The bots sit one level higher on the org chart. They plan, hand out work, remember things and ask questions. According to Fortune and ua.news, several features come together to make that work:

  • Memory that lasts weeks: The bot keeps the context of the whole project, not just of a single session. Without that, every handoff would start from scratch.
  • A teammate in Slack: The bot shows up in the Slack workplace chat like a colleague. Several people can work with the same bot, including over iMessage. When it gets stuck or needs a decision, it speaks up on its own instead of guessing or quietly stalling.
  • The right model for each phase: The bots are model-agnostic, so they aren't tied to any single provider's AI. For planning, design, computer use (where the AI operates a computer itself with mouse and keyboard) and other project steps, the bot picks a suitable model each time. For routine code it uses faster, cheaper models. Cognition also trains its own models. CEO Scott Wu says they're built for coding and code review and cost less than frontier models, the most capable models from the big labs. Devin combines them with models from other providers.
  • Its own budget: Each bot can get its own budget, measured against the results it delivers.

Wu sums up where this is heading:

"Self-driving software is, I think, the direction things are going. But of course, humans always decide where we're driving to."

Security: rules, not trust in the model

The bots run on their own cloud machines, not on developers' laptops. That way they never get near the passwords and access keys stored there. Admins decide which code repositories (where a project's source code lives) and tools each bot can access. Network controls are meant to stop data from leaking out without permission. That's supposed to hold even if a model falls for a prompt injection, meaning malicious instructions hidden in text it reads.

Code written by the bots has to pass the same automated tests as code written by humans. Teams decide which changes need a human to review them. Small, low-risk changes, like a new button color, can be approved and merged automatically, meaning they go straight into the main codebase. According to Wu, the safeguards rely on rules that companies set and monitor themselves, "rather than trusting that the models are smart enough to recognize a risky or malicious prompt."

The numbers, and how far you can trust them

Cognition says bot usage grew tenfold in two weeks among a few dozen early-access customers, including Decagon, Fireworks AI and Reducto. The dev.to author puts that in context: this is growth from a small base, not proof of quality at scale.

One game studio stands out even more. Its weekly pull requests reportedly went from about 50 to around 1,900. A pull request is a bundle of code changes submitted for review before it gets merged into the project. That's more than a thirtyfold increase, and a small human team oversaw the work. But the dev.to author points out that this measures volume, not quality. Cognition hasn't said how many of those changes were properly reviewed and merged, or how many caused bugs later.

The most solid data point doesn't come from Cognition at all. Citigroup CEO Jane Fraser said on the bank's earnings call on October 13 that 14,000 developers at Citi use Devin:

"We've brought work back in-house, which has cut our spending on technology contractors by 30 percent year over year."

That refers to the existing Devin product, though, not the new bots.

Metric Value
Valuation (September 2026) $48B after a $2B round
Valuation May 2026 / 2025 $26B / approx. $10B
Annualized revenue run rate over $1B (September), May: approx. $492M
Bot usage in early access tenfold in two weeks
Pull requests at the game studio approx. 50 → 1,900 per week
Devin users at Citi 14,000 developers

The annualized run rate takes current revenue and projects it over a full year. As Crypto Briefing notes, it shows current momentum and isn't revenue the company has actually booked for the year.

The backdrop: pressure from OpenAI and Anthropic

Cognition is in a tricky spot. The big model labs are both its suppliers and its competitors. Their progress makes Devin better, but improvements to OpenAI's Codex and Anthropic's Claude Code also give customers less reason to pay for Devin. Shortly before the bots launched, both labs shipped similar products. Around September 29, 2026, OpenAI introduced Dots, an always-on, proactive agent in Codex that directs teams of agents on development tasks. Anthropic followed around October 21 with Duet, a comparable product for Claude Code.

Cognition is betting on specialization. In its view, Dots and Duet are general-purpose agents that happen to code too, while Devin Bots are built from the ground up around a coding agent. Wu believes the product around the models matters more the smarter the models get. He pushes back on the idea that AI will eventually become so capable that "nothing else matters."

The road here hasn't been smooth. After the 2024 hype, SiliconANGLE says there were plenty of reports of Devin failing on complex codebases and producing buggy code. It also cost $500 per user per month at first. With Devin 2.0 in April 2025, Cognition cut the entry price to $20 and introduced a cloud development environment where several Devins can work in parallel. Looking back, that was an early version of today's bots. Another deal shows how hot this space is: in August, according to dev.to, SpaceX acquired Anysphere, the company behind the Cursor code editor, at a $60 billion valuation.

The catch: the review work just moves

The dev.to author raises the key objection. A bot directing a team of agents can produce a lot of code very quickly, and somebody has to read it. Writing code turns into reviewing code, setting rules and deciding what the bot is allowed to handle on its own. Small teams could really benefit. For large teams, there's an open question: how is code review supposed to work when the volume of incoming code goes up thirtyfold?

Cognition's answer is rules for automatic merging. So the author doesn't think the risk is a bot writing bad code. The bigger danger is a team that sets the rules too loosely and only notices the problem weeks later in production.

Our take

Devin Bots are the most concrete version yet of an idea the whole industry is moving toward: software development becomes something you manage rather than type out yourself. GitHub Copilot started out offering code suggestions you could accept or reject. A Devin Bot, by contrast, asks what it should optimize for, when it should send you updates, or whether you'd like a video of its test runs. For developers, that could change the job more than any single new model.

For now, though, the launch is still a promise. The growth figures come from a small test group, the 1,900 pull requests say nothing about quality, and the two most important details are missing: what a bot costs and when everyone can use it. Whether specialization can beat the general-purpose agents from OpenAI and Anthropic is also an open question. If you already use Codex or Claude Code, Dots and Duet give you similar features. A real comparison will only be possible once pricing is known.

Sources

#cognition#devin#ai agents#programming#coding
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.