← AUTOMATER NEWSROOM

Cognition Launches Devin Bots: Persistent Coordinators That Lead Teams of Devins for Days or Weeks

Cognition's Devin Bots turn Devin into a persistent coordinator that delegates goals to parallel Devin sessions, challenging OpenAI's Dots and Anthropic's Duet.

Diagram showing a Devin Bot coordinator delegating three parallel Devin sessions, merging their pull requests, and escalating one decision to a human Needs You inbox.Diagram showing a Devin Bot coordinator delegating three parallel Devin sessions, merging their pull requests, and escalating one decision to a human Needs You inbox.
Original illustration: a Devin Bot breaks a goal into tickets, runs Devin sessions in parallel, merges PRs that pass CI, and escalates only the decisions that need a human.

Cognition spent two and a half years teaching Devin to finish tasks. On November 17, 2026, it shipped the next layer: an agent that owns the goal above the tasks, and decides which tasks exist in the first place.

Devin Bots, announced today, are persistent agents with their own memory and tools that lead multiple Devin coding sessions in parallel. Give a Bot a goal — keep Android at parity with iOS, maintain an open-source repo, optimize server inference — and it investigates, breaks the goal into tickets, spins up Devin sessions to work them, reviews results, and decides what to pursue next. It runs on its own cloud machine, keeps context across sessions and days, and surfaces only the decisions that need a human. Fortune’s exclusive report frames it as Cognition’s answer to two rivals that beat it to the always-on-agent idea: OpenAI’s Dots, which runs agent teams through Codex, and Anthropic’s Duet, which does the same through Claude Code.

Diagram showing a Devin Bot coordinator delegating three parallel Devin sessions, merging their pull requests, and escalating one decision to a human Needs You inbox.Diagram showing a Devin Bot coordinator delegating three parallel Devin sessions, merging their pull requests, and escalating one decision to a human Needs You inbox.
Original illustration: a Devin Bot breaks a goal into tickets, runs Devin sessions in parallel, merges PRs that pass CI, and escalates only the decisions that need a human.

From tasks to responsibilities

The practical shift is in what you type. A task tells the agent what to build; a responsibility tells it what outcome to protect and leaves it to figure out how. Cognition’s examples: instead of porting a feature to Android, you say keep Android at parity with iOS. Instead of fixing one layout shift, you say keep the web app smooth as it changes. The Bot then has to notice problems you never flagged and interrupt you only when a decision genuinely needs your judgment. Defining success, building the feedback loop and setting access boundaries is now the engineering work; writing the ticket isn’t.

Side by side comparison of a task to complete versus a responsibility to own, with three example pairs and arrows between them.Side by side comparison of a task to complete versus a responsibility to own, with three example pairs and arrows between them.
Cognition's framing for Devin Bots: a task tells an agent what to build, while a responsibility tells it what outcome to protect and leaves it to decide how.

Three Bots, run inside Cognition

The launch post backs the pitch with internal case studies rather than a benchmark chart.

Bot Told to Result
Android Bot Autonomously build and product-manage the Android app (one engineer, Sept. 28) 1,134 PRs merged, 1,217 tickets closed, 344,651 lines of Kotlin; shipped to Google Play open testing in 44 days after 183 human messages
Layout Bot Keep the web app’s UI from shifting; watch a check every 3 hours Rebuilds and replays production load every 3 hours; 159 PRs merged in 3 weeks; layout lag cut 70%
~20 Bots plus a coordinator Run a mobile game studio end-to-end Weekly PRs rose from ~50 to 1,900; studio now ships weekly with one human, who then joined Cognition full time

The Android case shows the shift clearest. Samyok, the engineer who created the Bot, didn’t write 1,217 tickets — he gave one directive and let the Bot discover blockers, file work, start a Devin session per ticket, and merge whatever passed CI and Devin Review. On 22 of the 44 days, he sent it nothing. Most of his 183 messages were decisions the Bot specifically needed, like flagging a PR for security review, not status pulls.

Cognition's Devin blog post titled Introducing Devin Bots, from tasks to outcomes, published Nov. 17, 2026.
Cognition's Devin blog, Introducing Devin Bots: from tasks to outcomes (Nov. 17, 2026), announcing persistent coordinator agents that lead teams of Devin sessions. · Original source

The mechanics behind the pitch

devin.ai/bots describes the plumbing worth checking before handing a Bot real responsibility. Bots watch triggers — a Slack channel, a schedule, a Linear board — and spin up individual Devin sessions on separate cloud machines when work needs doing. Each session still runs the standard Devin pipeline: codebase understanding, a cloud computer per agent, automated testing, Devin Review before a PR opens. Decisions land in a Needs You inbox rather than blocking the Bot. Templates ship for common jobs including Software Factory, Android Engineer and QA Engineer. Enterprise controls include per-session VMs, RBAC/SSO/SCIM/OIDC, private networking, audit logs and zero data retention enforced with model providers. Cognition folded in Poke-maker Interaction Company, acquired in July, for how Bots message users across Slack, iMessage and group chats.

Why bet on the layer above the model

CEO Scott Wu told Fortune his bet is that product matters more as models get smarter, not less — a direct answer to the argument that frontier models will eventually make surrounding tooling irrelevant. Devin Bots lands five weeks after Cognition’s October benchmark push with Devin Fusion, which paired Opus 5.5 with GPT-6 Luna to claim a leading coding-agent benchmark score at lower cost — a reminder that Cognition is shipping on the model-routing layer and the product layer at the same time, not choosing between them. Wu’s stated edge here: a Bot isn’t tied to one lab’s models, routing planning, design, computer use and routine code to whichever model fits, mixing Cognition’s own trained models with outside labs’, with each Bot carrying its own budget measured against results. On safety, Wu described the controls as resting on rules a company sets and monitors — which repos and tools a Bot can reach, what tier of change auto-merges, network controls meant to block unauthorized data transfers even if a model follows instructions hidden in content it reads — rather than hoping models recognize a risky prompt themselves. That’s a meaningfully different trust model than self-policing, and it’s the piece worth copying regardless of which coordinator you run: tests plus review, human sign-off for anything not pre-approved.

Devin already has scale behind it: Citigroup CEO Jane Fraser said on the bank’s Oct. 13 earnings call that 14,000 of its engineers use Devin and that in-sourcing development with it cut tech vendor spend 30% year-over-year. Cognition says usage of Bots specifically, across dozens of early-access customers including Decagon, Fireworks AI and Reducto, grew tenfold in two weeks.

What to do with this

If you’re already running Devin, Codex or Claude Code, Devin Bots is less a new tool to evaluate than a new unit of delegation to decide whether you trust. Before handing one a responsibility instead of a ticket, write a finish line the transcript can prove: write down what success means for the scope given, how it will evaluate its own work, and which systems it can touch without asking. Cognition’s post says as much — the Android Bot worked because Samyok set direction and constraints before it started filing tickets. Start narrow and cheaply verifiable, like a flaky-test triager or a dependency-bump Bot, before handing one a goal as broad as an entire app, and keep the fleet discipline — one place that tracks what’s running and what’s waiting on you — that any multi-agent setup needs once an agent can act for days unsupervised.

Devin Bots are live now at devin.ai/bots for existing Devin customers; Cognition has not published standalone pricing beyond its existing Devin plans.

Sources

Sources

  1. Introducing Devin Bots: from tasks to outcomes | Devin
  2. Devin Bots | Devin
  3. Exclusive: Cognition launches Devin Bots to compete with OpenAI's Dots and Anthropic's Duet | Fortune