Devin Bots Explained: Cognition’s AI Manager That Runs Coding Teams

On 17 November 2026, Cognition gave its famous AI software engineer a boss. Devin Bots are persistent manager agents that supervise teams of Devin coding agents on projects lasting days or even weeks — keeping the memory, breaking the work into pieces, handing it out, checking the results, and only bothering a human when a real decision is needed. Cognition’s example says it all: hand a Bot the job of building an Android version of your iPhone app, and it will keep the two apps in sync as the iPhone version keeps changing, week after week.

This guide explains what Devin Bots actually are, how they differ from the Devin you already know, what they cost, how they compare with Anthropic’s Duet and OpenAI’s Dots, and whether your team should try to get access now — or wait for the wider release.

Table of Contents

Devin Bots — software developers reviewing code together on a large screen, the team workflow Devin Bots is built to manage
Devin Bots are built for the way real teams already work: several people, several screens, and a long list of jobs that have to stay in sync for weeks. The Bot sits above the coding agents the way a manager sits above a team.

What Are Devin Bots?

Devin Bots are a new layer on top of Devin, Cognition’s autonomous AI software engineer. A normal Devin session takes one task — fix this bug, build this feature, write these tests — and works on it in its own cloud computer until it can open a pull request. A Devin Bot does something bigger: it owns a goal end to end. You give it a responsibility, such as “keep our Android app in step with the iPhone app” or “migrate this service to the new framework”, and it keeps working on that responsibility for days or weeks, running team after team of Devin sessions underneath it.

The difference is persistence. A single Devin session starts fresh, does a job, and finishes. A Devin Bot remembers. It carries the context of the whole project across sessions — what was tried, what worked, what failed, what is still open — so the work does not restart from zero every morning. That memory is the feature that makes multi-week work possible at all. Without it, every handoff would lose the thread, and a long project would collapse into a pile of disconnected tasks.

Each Bot also has its own tools and its own budget. Cognition says a Bot picks the right model for each stage of the work — a cheaper, faster model for routine code, a stronger model for planning and design — and each Bot can be given a spending budget that is measured against the results it delivers. In other words, a Bot is not just a smarter prompt. It is a managed worker with a job description, a memory, a team, and a budget.

Cognition launched Devin Bots on 17 November 2026, and the timing was no accident. OpenAI shipped Dots, its always-on proactive agent built on Codex, a few weeks earlier. Anthropic shipped Duet, a similar always-on agent tied to Claude Code, around the same time. All three companies are racing to answer the same question: what happens when an AI agent stops being a tool you open for one task and starts being a colleague that keeps working while you do something else? Devin Bots are Cognition’s answer — and, given that Devin was the first autonomous coding agent to become famous back in March 2024, the company is betting its head start matters.

How Devin Bots Work

A Devin Bot works less like a chatbot and more like a junior engineering manager who never sleeps. You hand it a goal. It investigates the problem, breaks the work into pieces, assigns those pieces to individual Devin sessions running in parallel, evaluates what comes back, and decides what to try next. Only the decisions that genuinely need human judgement reach you, in a “Needs You” inbox. Everything else, the Bot handles on its own.

A memory that lasts for weeks, not one session

The core of a Devin Bot is its memory. Traditional coding agents are forgetful by design: each session is a fresh conversation, and anything learned in one session has to be re-explained in the next. A Devin Bot keeps a running record of the whole project — the architecture, the decisions, the dead ends, the open questions — and carries it from one Devin session to the next. That is what lets it run a project for weeks instead of hours. It also means the Bot gets better at the project as it goes, because it is not re-learning the codebase every time it starts a new piece of work.

One Bot, many Devin sessions

Underneath every Bot sits a team of ordinary Devin agents. The Bot plans the work, splits it into discrete tasks, and delegates them to Devin sessions that run in parallel on separate cloud computers. Each session works in its own virtual machine, with its own shell, editor and browser, and returns its work as a pull request — the same way a human developer would. The Bot then reviews the results, merges what is good, sends back what is not, and re-plans. This hierarchy — one manager, many workers — is the whole idea. A single agent can write a function. A managed team of agents can rebuild a feature, migrate a service, or keep two apps in sync for a month.

The “Needs You” inbox: humans decide, Bots do the rest

Cognition is careful to say that humans stay in charge of the destination and the priorities. The mechanism is a “Needs You” inbox: the Bot brings only the decisions that need your judgement to you, and gets on with the rest. Teams also decide which changes require human review. Low-risk edits — changing a button’s colour, for example — can be configured for automatic approval and merging. Bigger changes go through the same tests and reviews as human-written code before anything lands. The point is control without babysitting: you set the rules once, and the Bot works inside them.

Working where your team already works

A Devin Bot does not live in a separate dashboard you have to remember to open. It joins your team’s Slack workspace as its own team member, and several people can work with the same Bot — including over iMessage. You coordinate with it the way you would coordinate with a colleague: in the chat thread, in plain language. That matters more than it sounds. The biggest reason earlier coding agents stayed a side tool is that using them meant leaving the flow of work. A Bot that lives in Slack meets the team where the decisions already happen.

Model-agnostic, with its own budget

Devin Bots are model-agnostic. A Bot selects suitable models for planning, design, computer use and other parts of a project, using faster, cheaper ones for routine code. Cognition also trains its own models — its SWE-2 family is designed for software engineering and code review at lower cost than frontier models — and Devin combines those with models from outside labs, including OpenAI, Anthropic and Google, plus open-source models. Each Bot can be given its own budget, measured against the results it delivers, so a team can cap what a Bot is allowed to spend on a project before it starts.

Running on separate cloud computers

Bots run on separate cloud computers, away from engineers’ laptops, so they never touch the passwords and keys stored on a developer’s machine. That separation is a genuine security feature, not a footnote: an agent that can read code, run commands and open pull requests is powerful, and the blast radius of a mistake is much smaller when it runs in an isolated cloud workspace than on a laptop full of credentials.

Devin Bots vs Devin: What Is Actually New?

If you already use Devin, the honest question is what a Bot adds. The answer is scope and staying power. A Devin session is a task worker. A Devin Bot is a project owner. Here is the difference in plain terms:

  • Time horizon: a Devin session handles a task that a human could do in about three hours. A Devin Bot handles a project that takes days or weeks.
  • Memory: a Devin session starts fresh. A Devin Bot remembers the whole project — decisions, failures, open questions — across every session it runs.
  • Delegation: a Devin session does the work itself. A Devin Bot plans the work and delegates it to many Devin sessions running in parallel, then reviews and integrates the results.
  • Interaction: you talk to Devin through tickets, the web app or the command line. A Devin Bot joins Slack as a team member, and a whole team can work with the same Bot.
  • Authority: a Devin session needs you to check each result. A Devin Bot can be configured to auto-approve and merge low-risk changes, and only escalates real decisions to your “Needs You” inbox.

None of this replaces Devin. The Bot is a layer above it: the same coding agents, the same cloud workspaces, the same pull requests — organised by a manager that never forgets the plan. If your work is a stream of small tickets, plain Devin is still the right tool. If your work is a project with a beginning, a middle and an end that lasts for weeks, that is what a Bot is for.

Devin Bots — developers collaborating around code on several screens, the kind of teamwork Devin Bots coordinates across weeks
One agent can write a function. A Bot coordinating a team of agents can rebuild a feature or keep two apps in sync for a month — that is the jump Devin Bots are trying to make.

What Can Devin Bots Actually Do?

Cognition’s launch examples are deliberately ordinary, because ordinary is where the money is. The headline example is building an Android version of an iPhone app and keeping it updated as the iPhone version changes — the kind of grinding, never-finished maintenance work that eats engineering time. Beyond that, the use cases Cognition describes fall into four groups:

  • Cross-platform development: two codebases that must stay synchronised over weeks of parallel change — iOS and Android, web and mobile, or a product and its white-label twin.
  • Large refactoring projects: coordinated changes across dozens of modules, where one change in one file breaks three others unless somebody is tracking the whole picture.
  • Feature rollouts: multi-stage work with testing, review and deployment cycles, managed by a single Bot from first commit to release.
  • Migrations and maintenance: moving a service to a new framework, paying down technical debt, or clearing a backlog of chores that never quite reaches the top of the sprint board.

What a Bot will not do — at least not yet — is replace the judgement calls. Architecture decisions, release decisions, spending decisions and anything with real business risk still land in the “Needs You” inbox. Cognition’s chief executive, Scott Wu, put it this way to Fortune: self-driving software is the direction of the industry, but humans keep control of the destination and the priorities. That framing is worth remembering when the marketing gets loud. A Bot is a very capable employee, not an autopilot.

Devin Bots Pricing: What Will a Bot Cost You?

Here is the honest answer: Cognition has not published a separate price for Devin Bots. The company says Bots run on Devin — there is no separate “Bot plan” on the price list — and no public figure exists for what a Bot costs or how the per-Bot budget is charged. If you see a website quoting a Devin Bots price today, it is guessing.

What we do know is Devin’s existing pricing, because that is the platform a Bot runs on. Devin has a free plan with a light agent quota, a Pro plan at $20 a month with daily and weekly quotas covering Devin sessions, the command line and the desktop app, a Max plan at $200 a month with a much larger weekly allowance, and a Teams plan starting at $80 a month plus $40 a month per full seat. Usage beyond the included quota is metered by the actions Devin takes and the virtual-machine time it uses, and extra usage is bought at API pricing.

That metering matters for Bots: a Bot running for weeks, delegating to many sessions, will burn through usage far faster than a single developer running occasional sessions. Until Cognition publishes Bot pricing, the safe assumption is that a Bot is a Teams-scale purchase, not a $20 experiment — and that the per-Bot budget control is there precisely because the bill can grow.

There is one more cost to budget for: review time. A game studio in the early-access programme went from about 50 pull requests a week to 1,900. Somebody still has to look at those pull requests, or at least set the rules for which ones merge automatically. The cheapest Bot in the world is expensive if it floods your team with code nobody has time to review.

The Numbers Cognition Is Bragging About

Cognition launched Devin Bots with a set of numbers designed to impress, and every one of them comes from Cognition itself — there are no independent benchmarks and no third-party audit behind them. Read them as the company’s claims, not as tested facts:

  • Tenfold usage growth: among the early-access customers, Bot usage grew tenfold within two weeks of availability, according to Cognition.
  • 50 to 1,900 pull requests: one game studio in the programme saw weekly pull requests climb from about 50 to 1,900, with a small human team overseeing the expanded output.
  • 14,000 engineers at Citigroup: the bank already has 14,000 engineers using the original Devin, and the Bots tier is pitched as the way to orchestrate squads of those agents on longer initiatives.
  • $48 billion valuation: Cognition was valued at $48 billion after a $2 billion funding round in September 2026.
  • $1 billion annualised revenue: the company passed a $1 billion annualised revenue run rate in September 2026, according to Fortune.
  • Early customers: Decagon, Fireworks AI and Reducto are named as early-access customers.

The 1,900-pull-request figure deserves a second look. It proves volume, not quality. A Bot that opens 1,900 pull requests a week is only useful if most of them are good — and Cognition has published no merge rate, no defect rate and no figure for how much of that code survived review. The tenfold usage growth is a stronger signal, because it measures what customers actually chose to do, not what the product produced. Either way, the numbers tell you the direction of travel: teams that try Bots use them a lot more, very quickly. Whether that is excitement or dependence is the question every engineering manager should ask before signing up.

Devin Bots — a developer coding late at night on a laptop, the long hours Devin Bots are designed to take over on multi-week projects
A Bot does not get tired at 2 a.m. — but the code it writes at 2 a.m. still has to be reviewed. The volume numbers are impressive; the quality numbers have not been published yet.

Devin Bots vs Anthropic Duet vs OpenAI Dots

Devin Bots do not arrive in a vacuum. Within a few weeks, three of the biggest names in AI shipped three versions of the same idea: an always-on agent that keeps working on your behalf. Here is how they line up:

Feature Devin Bots (Cognition) Duet (Anthropic) Dots (OpenAI)
Launched 17 November 2026 October 2026 October 2026
Built on Devin coding agents Claude Code Codex
Core idea A persistent manager agent that runs teams of Devin sessions for weeks An always-on agent tied to Claude Code An always-on proactive agent inside Codex
Memory Project memory across all sessions Carries context across sessions Carries context across sessions
Team chat Slack and iMessage, as a team member Works inside Claude Code workflows Works inside Codex workflows
Human control “Needs You” inbox; configurable auto-merge for low-risk edits Human review gates Human review gates
Pricing Not published separately; runs on Devin plans from $20/month Tied to Claude plans Tied to ChatGPT/Codex plans
Availability Early access, dozens of customers Rolling out Rolling out

The real difference is focus. Duet and Dots are general always-on agents built on top of each company’s coding tool. Devin Bots are narrower and deeper: they only do software work, but they do it with a management layer — planning, delegation, budgets, a team that lives in Slack — that the other two do not advertise. Cognition’s bet is that software teams want a manager, not just a faster coder. Anthropic and OpenAI are betting that a general agent attached to the tool developers already use will win instead. Both bets can be partly right, and the winner will be decided by which one engineering managers actually trust with a production codebase.

If you want the full picture on the rival products, read our guide to Anthropic’s Claude Duet — the always-on agent Devin Bots will be measured against — and our explainer on Cognition’s own SWE-2 Flash model, the coding model that powers much of Devin’s work. For the wider comparison of coding tools, our GitHub Copilot Fusion guide covers the other major player in the same race.

Who Should Use Devin Bots?

Devin Bots are not for everyone, and Cognition’s own launch story tells you who they are for. The early-access customers are companies with large codebases, long projects and engineering managers who are already drowning in coordination work. If that sounds like your team, a Bot may be worth the wait. If it does not, plain Devin — or a rival coding assistant — is the better fit for now.

A Devin Bot is probably worth trying if:

  • Your team maintains two or more versions of the same product — iOS and Android, web and mobile — and keeping them in sync is a permanent tax on your roadmap.
  • You have a migration, refactor or framework upgrade that has been “next quarter” for a year, because no sprint ever has room for it.
  • Your backlog is full of well-defined chores — dependency updates, test coverage, small features — that a supervised agent could clear while your engineers do the hard work.
  • You already use Devin and trust it with pull requests, and you want the same quality of work sustained over a multi-week project instead of one ticket at a time.

A Devin Bot is probably not for you yet if:

  • You are a solo developer or a two-person team. A Bot is a manager; a team of one does not need a manager.
  • Your codebase is small enough that one person can hold it in their head. The Bot’s memory is valuable precisely when the project is too big for any one person to remember.
  • You cannot spare review time. A Bot that produces 1,900 pull requests a week is a liability if nobody reads them.
  • You need predictable pricing today. With no published Bot price and metered usage underneath, the bill for a multi-week project is an open question until Cognition says otherwise.

The Risks: What to Watch Before You Hand a Bot the Keys

The launch coverage is enthusiastic, and most of it is written from Cognition’s own announcements. Before you hand a Bot access to your repositories, here are the risks the launch material does not dwell on:

Authority is the real trade-off. A Bot acts inside your repositories and tools with whatever permissions you grant it. Scope each Bot narrowly — one project, one set of repositories, one spending budget — and keep release and spending decisions in the “Needs You” queue. A Bot with broad access and auto-merge turned on is a deployment pipeline with no human in the loop, and that is a choice, not a default.

Volume is not quality. Cognition has published no benchmarks for Devin Bots and no quality figures — no merge rate, no defect rate, no comparison against human-written code. The 1,900-pull-request story is a volume metric. Until independent teams publish their own results, treat every output figure as a claim.

Costs are hard to predict. Devin usage is metered by actions taken and virtual-machine time, on top of subscription quotas. A Bot that delegates to dozens of sessions for weeks can run up usage fast, and the per-Bot budget is a cap, not a price. Ask for the pricing before you commit to a project, and set the budget low on the first one.

Early access means rough edges. Bots are in early access with dozens of customers. There is no public date for wider access, no public documentation of failure modes, and no track record of how a Bot behaves when a project goes wrong — when the plan is bad, when the tests are wrong, or when two Bots working on related projects step on each other. (Cognition says multiple Bots can coordinate with each other, but has shown no detail of how that works in practice.)

Your team still owns the result. If a Bot merges a bad change, the incident is yours. The “Needs You” inbox only helps if somebody is watching it. The teams that will get the most from Devin Bots are the ones that treat the Bot like a talented new hire: give it clear goals, check its work early and often, and tighten the leash before you loosen it.

Devin Bots — a team meeting around laptops to plan work, the kind of planning a Devin Bot takes over and keeps running for weeks
The teams that get the most from a Bot will treat it like a talented new hire: clear goals, checked work, and a tight leash at first.

How to Get Access to Devin Bots

Devin Bots are available now at devin.ai/bots, but only in early access. Cognition has not announced a public rollout date, and the launch material names dozens of early-access customers rather than an open waiting list. If you want in, the practical steps are:

  1. Start with Devin itself. Bots run on Devin, and Cognition’s early-access customers are existing Devin users. Set up a Devin account — the free plan is enough to learn the tool — and use it on real tickets before you ask for a Bot.
  2. Pick one project, not five. Choose a well-defined, multi-week job with a clear definition of done — an Android port, a migration, a refactor. A Bot needs a goal it can own; “help with everything” is not a goal.
  3. Set the guardrails first. Decide which repositories the Bot can touch, which changes need human review, which low-risk edits can auto-merge, and what its spending budget is — before it writes its first line of code.
  4. Apply through the official page. Go to devin.ai/bots and follow Cognition’s early-access process. Do not trust third-party sites offering “Devin Bots access” — the only official route is Cognition itself, at cognition.com.

Frequently Asked Questions

What are Devin Bots?

Devin Bots are persistent manager agents from Cognition, launched on 17 November 2026. Each Bot owns a goal end to end, keeps the memory of a whole project for weeks, and delegates the work to teams of Devin coding agents running in parallel — only escalating decisions that need human judgement to a “Needs You” inbox.

How are Devin Bots different from Devin?

Devin handles one task at a time in a single session. A Devin Bot handles a whole project over days or weeks: it remembers the project, plans the work, splits it across many Devin sessions, reviews the results, and keeps going until the goal is done. Think of Devin as a coder and a Devin Bot as the coder’s manager.

How much do Devin Bots cost?

Cognition has not published a separate price for Devin Bots. The company says Bots run on Devin, so they sit on top of Devin’s existing plans: free, Pro at $20 a month, Max at $200 a month, and Teams from $80 a month plus $40 per full seat, with extra usage metered. Each Bot can also be given its own spending budget. Expect a Bot to be a Teams-scale purchase, not a $20 experiment.

Are Devin Bots available to everyone?

Not yet. Devin Bots launched in early access on 17 November 2026 with dozens of customers, including Decagon, Fireworks AI and Reducto. Cognition has not announced a date for general availability. You can apply through the official page at devin.ai/bots.

Do Devin Bots replace software engineers?

No — and Cognition does not claim they do. A Bot handles the planning, delegation and routine coding of a long project, but humans set the goals, make the real decisions through the “Needs You” inbox, and review the code. The honest framing is that a Bot removes the coordination tax on long projects; the engineering judgement stays with people.

How do Devin Bots compare with Anthropic Duet and OpenAI Dots?

All three are always-on agents that keep working over days or weeks. Duet is tied to Claude Code and Dots is tied to Codex; both are general agents. Devin Bots are narrower — software work only — but deeper, with a management layer (planning, delegation, per-Bot budgets, Slack team membership) that the other two do not advertise. Cognition’s bet is that software teams want a manager, not just a faster coder.

Can a Devin Bot merge code without a human checking it?

Yes, if you configure it that way — teams can set low-risk edits, such as changing a button’s colour, to merge automatically. Bigger changes go through the same tests and reviews as human-written code. Our advice: keep auto-merge off until you have seen a Bot’s work on your own codebase for a few weeks.

The Bottom Line

Devin Bots are the most important launch in the coding-agent race since Devin itself. The idea is simple and, for once, the marketing is honest about what changed: the bottleneck in AI-assisted coding was never writing a single function — it was keeping a whole project coherent for weeks while dozens of small jobs piled up.

A persistent manager agent with memory, a team underneath it and a budget above it is a genuine answer to that problem, and the early numbers — tenfold usage growth in two weeks, a game studio jumping from 50 to 1,900 weekly pull requests, 14,000 engineers at Citigroup already on Devin — suggest teams that try Bots do not go back to one-task-at-a-time agents.

But the launch also leaves the three questions that matter most unanswered: what a Bot costs, how good its code actually is, and when the rest of us can use one. There are no published benchmarks, no merge-rate figures and no public price. Until those appear, the sensible move is to start with Devin itself, learn what its agents do well and badly on your own code, and apply for Bot access with one well-defined project and tight guardrails. The manager era of coding agents has started. The teams that win in it will be the ones that manage their managers.

Devin Bots launched on 17 November 2026 and are in early access. Pricing, availability and performance figures in this article are as reported by Cognition and Fortune at launch; check devin.ai/bots for the latest status before you buy.

Leave a Comment

Your email address will not be published. Required fields are marked *