Cognition has introduced Devin Bots, persistent agents that own a goal and coordinate several Devin coding sessions at once. Announced on November 17, the launch pushes the company's autonomous software engineer beyond single assignments and toward always-on teams that watch for work, divide it up and bring in people only when a decision is needed.
From Single Tasks to Managed Teams
Devin has been able to take on one engineering assignment at a time. Devin Bots add a management layer on top. Cognition describes them as persistent, goal-owning agents that keep memory across sessions, inspect repositories and production telemetry, break long-running goals into parallel Devin tasks, and follow up when work stalls.
In practice, a Bot can triage overnight incidents, review pull requests, run post-deploy checks for regressions and send a morning briefing that lists only the architectural or product choices that still need a human. The shift is from supervising each code change to supervising outcomes. That is a meaningful change for engineering leads, who would otherwise spend their time reviewing every session.
How the Coordination Works
A Bot splits a request into parallel workstreams, and each Devin session runs on its own cloud machine. Bots can also hand work to one another. A QA Bot might flag a bug, bring in a developer Bot to fix it and then test the result, with the whole loop running without a person in the middle.
Bots inherit a team's existing workspace integrations, repository permissions, CI/CD hooks and incident channels, so they do not need a separate setup, according to coverage of the launch. Triggers are the other key piece. A Bot can watch a Slack channel for customer issues, run a nightly CI health check or monitor a Linear board for blockers. People can message a Bot by text, in Slack, by phone call or in a group chat.
Early Results and Availability
Cognition's own Devin blog, as cited in coverage, says internal Bots merged roughly 1,600 pull requests in a single week. That work covered open-source maintenance, release management and inference optimization. The company also says an Android Bot was assigned the Devin Android app in late September and built it without dedicated Android staff. These are company-reported figures and have not been independently verified.
Cognition has not announced standalone pricing for Bots. They build on Devin Apps and the upgraded Devin Teams plan. The product page at devin.ai/bots lists enterprise controls including SOC 2, ISO 27001 and FedRAMP compliance, per-session virtual machines, single sign-on with role-based access, audit logs and optional private networking. A template library offers starting points such as a Software Factory Bot, a QA Engineer Bot and a Product Engineer Bot.
Cognition's momentum is part of the context. The company said in September that it was on track for $1 billion in annualized revenue, and it sells to large enterprises including Citigroup and Nvidia, according to reporting on its business. For teams evaluating Bots, the practical starting point is the trigger layer. A Bot that watches one Slack channel for customer issues has a much smaller blast radius than one that can start sessions from every repository and board. Engineering leaders should list which triggers exist, who can create them and which credentials each Bot can use before rollout.
Why It Matters
Agent products are moving from tools that answer requests toward managers that pursue goals over days or weeks. That changes who reviews the work. Instead of checking each change as it appears, teams will need to set boundaries, define which decisions a Bot may make alone and audit what it did afterward.
The security stakes rise with persistence. A long-lived agent with its own secrets and broad repository access has a larger attack surface than a session that ends when its task does. Companies evaluating Bots should start with narrow triggers, read-only access where possible, mandatory human review before merges and regular checks of audit logs.
The launch also puts Cognition in direct competition with other coding agent vendors that are adding multi-agent features. Buyers will likely judge these products on results from real pull requests and on incident response, rather than on headline counts alone. Independent measurements of how often Bots' merged work holds up in production will be the number to watch.
