Cognition's Devin Bots Now Run Parallel Coding Sessions, Hitting 2,000 Pull Requests Weekly

Cognition's new Devin Bots act as managers that direct fleets of coding agents, with beta testers pushing up to 2,000 PRs weekly.

·
·
·
  • Cognition launched Devin Bots, supervisory agents that coordinate teams of Devin coding agents in parallel.
  • Beta testers reportedly merge 4.5x more PRs per message, with some engineers hitting nearly 2,000 PRs weekly.
  • Bots interact via Slack, text, phone calls, group chats, and scheduled triggers watching CI and Linear.
  • Launch templates include Software Factory, Android Engineer, QA Engineer, Latency Optimizer, Product Engineer, Chief of Staff.
  • Multi-agent handoffs built in: QA Bot finds a bug, hands to Developer Bot, retests the result.
  • Ships with SOC 2, ISO 27001, FedRAMP, per-session VMs, zero data retention with model providers.

Cognition has introduced Devin Bots, persistent software coordinators that divide broad goals into tasks, launch multiple Devin coding sessions, track their progress, and escalate decisions that require human approval.

A Bot can own an outcome such as keeping continuous integration healthy, clearing a ticket queue, or testing incoming pull requests. It can receive work through Slack, Linear, alerts, schedules, text messages, and phone calls, then continue operating while developers are offline. The model could increase engineering throughput, provided teams can review and validate the resulting code.

One Bot, many Devin sessions

Each Bot adds a coordination layer to Devin, Cognition’s autonomous coding environment. A Devin session can use a code editor, browser, shell, and isolated virtual machine to inspect a repository, modify code, run tests, and prepare a pull request.

  1. Collect work: The Bot watches configured channels, tickets, alerts, schedules, and CI results.
  2. Plan execution: It breaks an objective into smaller assignments and selects the appropriate workflow for each one.
  3. Run in parallel: It launches separate Devin sessions so multiple assignments can proceed concurrently.
  4. Coordinate handoffs: One Bot can pass work to another, such as sending a QA finding to a Developer Bot for repair.
  5. Request approval: It contacts a developer when a task requires judgment, access, or permission to proceed.

Developers can follow the work in their existing communication channels. Cognition’s examples include a QA Bot that identifies a defect, delegates the fix, and reruns the tests, plus a Chief of Staff Bot that distributes requests and tracks them through completion.

The beta numbers need denominators

Cognition reports substantial throughput gains among beta users, although the company has not released enough supporting data to assess code quality, repository complexity, or the representativeness of the results.

Metric Reported result What it measures
Merged pull requests per message 4.5 times higher with Bots than with direct Devin sessions How much accepted work followed each user interaction
Peak weekly pull-request volume Some engineers rose from about 50 to nearly 2,000 The upper end of company-reported output, rather than a typical result

Neither metric establishes whether the pull requests were small, substantive, secure, or durable after deployment. Cognition has not published sample sizes, median results, defect rates, rollback rates, review time, compute costs, or independent replication. Those details will determine whether higher concurrency produces useful software or transfers work into review queues.

Six jobs arrive prepackaged

Cognition supplies templates that package triggers, instructions, schedules, and handoffs for common engineering roles, reducing the amount of orchestration that teams must configure themselves.

  • Software Factory: Finds bugs, cleanup opportunities, and refactoring work within a product area, then opens a small pull request for each change. Feedback from merged and closed requests informs later work.
  • Android Engineer: Takes a Kotlin ticket through implementation and produces a build for internal testers.
  • Latency Optimizer: Searches for slow endpoints and database queries, then opens weekly pull requests with before-and-after measurements.
  • QA Engineer: Tests each pull request in a browser and attaches video showing successful and failed checks.
  • Chief of Staff: Posts updates, delegates incoming work, and tracks assignments to completion.
  • Product Engineer: Converts a ticket into a tested pull request and manages the review process.

Cognition describes these templates as available at launch. Its announcement does not specify Bot-specific pricing, usage limits, or rollout eligibility.

Enterprise controls shape the pitch

The launch extends Cognition’s effort to sell Devin as an enterprise engineering platform. The company has also released Devin Apps for hosted internal tools and announced deployments or partnerships involving Mercedes-Benz and McKinsey.

Cognition says Bots include persistent memory, scheduled triggers, repository context, and an isolated virtual machine for each session. The company also lists role-based access control, single sign-on, SOC 2, ISO 27001, and zero data retention with model providers, which means those providers are instructed not to retain customer prompts and outputs.

The platform is listed on the FedRAMP Marketplace as FedRAMP Class D High In Process, according to Cognition. That designation indicates authorization work is underway; final FedRAMP authorization remains pending.

Review capacity sets the ceiling

Managed orchestration is the product’s clearest distinction from multi-agent frameworks that require teams to build their own routing, memory, scheduling, and monitoring systems. Devin Bots package those functions into a product connected to Cognition’s existing coding agents.

Teams evaluating the system will need controls that match its concurrency. Useful safeguards include protected branches, mandatory tests, narrow credentials, repository boundaries, human approval for sensitive changes, complete audit logs, spending limits, and automatic suspension after repeated failures.

The practical measure of success is validated throughput: changes that pass tests, survive review, reach production, and remain stable. Cognition’s beta figures suggest that Bots can increase parallel work, while questions about cost, security, maintenance burden, and reviewer capacity remain open.

Trending
  • No trending articles

Comments

avatar

Next Reads