Cognition launched Devin Bots on November 17 — a persistent agent that manages teams of Devin coding agents on multi-week software projects. You give it an outcome. It plans the work, spawns agents, delegates tasks, and runs until the goal is delivered. The headline number from Cognition’s own internal deployment: 183 human messages produced 344,651 lines of Kotlin and 1,134 merged pull requests over 44 days. That is not a benchmark. That is an Android app.
From Tasks to Outcomes
The distinction Cognition is drawing matters. Devin, the original autonomous coding agent, works on discrete tasks — you describe a job, it completes it. Devin Bots operates at a higher level. Each Bot carries persistent memory across the entire project, breaks the goal into work units, and delegates those units to individual Devin instances running in parallel. When one finishes, the Bot reviews the output, adjusts the plan, and assigns the next piece.
Cognition’s framing is blunt: “The next phase of AI coding is management, not autocomplete. The question is no longer ‘can an agent write this function’ — it’s ‘can a team of agents run this project while humans do something else.'”
That is a meaningful shift. Single-agent coding assistants required constant developer attention. Devin Bots is designed to disappear for days and resurface with pull requests.
How It Works
Each Bot is model-agnostic — it selects the appropriate model for each stage of a project: heavy reasoning models for planning and architecture decisions, faster and cheaper models for boilerplate and routine code. The whole system runs off-laptop, which means it never touches local credentials or API keys. Developers can assign each Bot a budget and measure it against delivered results.
The architecture follows a hierarchical multi-agent pattern, but Cognition’s execution is what has early customers excited. Devin’s existing integrations, permissions, and tool access carry over to Bots automatically, eliminating the setup friction that makes most multi-agent frameworks painful in practice.
The Numbers Are Real
Beta results are specific enough to take seriously:
- A game studio moved from roughly 50 pull requests per week to 1,900 — with a small human oversight team.
- Reducto, an early-access customer, made Devin Oncall the first responder to most engineering alerts. On-call engineers went from 3–5 pages per week to almost none.
- Across early-access customers, Bot usage grew tenfold in two weeks.
- Beta testers report 4.5x more PRs per message using Devin Bots.
Reducto’s feedback is direct: “This is by far the most effective persistent agent I’ve used. It defaults to real engineering work, instead of just summarizing my calendar and emails.”
That last sentence is a quiet jab at the broader persistent-agent category, where “proactive assistance” often means managing your to-do list. Devin Bots has a narrower, more demanding brief: ship code.
The Competitive Moment
The timing is not coincidental. OpenAI released Dots — its persistent agent product — on September 29. Anthropic opened research access to Claude Duet on October 21. Within seven weeks, all three major AI coding players launched persistent agent offerings.
Dots and Duet are positioned as general-purpose proactive assistants. They handle email, meetings, research, and code. Cognition’s bet is that developers doing serious production work want something purpose-built: an agent that owns a software project and treats pull requests as its primary unit of output, not calendar summaries.
This positioning plays to Cognition’s core strength. The company was founded by competitive programmers, raised $2 billion at a $48 billion valuation in September 2026, and counts Mercedes-Benz, NASA, Goldman Sachs, and Citi among its enterprise customers. Devin Bots is the product of a team that has been thinking about automated software engineering — not general assistance — since day one.
Availability
Devin Bots is in early access. Current customers include Decagon, Fireworks AI, and Reducto. Cognition has not announced public pricing or a general availability date. Access is through Devin’s existing chat interface: assign an outcome to a Bot and it takes it from there.
The managed-agent-team pattern Devin Bots represents is coming regardless of which company wins this round. Anthropic’s 2026 Agentic Coding Trends Report found that 72% of enterprise AI projects now involve multi-agent architectures, up from 23% in 2024. The infrastructure is arriving fast. The question is who builds the management layer developers actually trust with production systems.

