How Cognition builds Devin using Devin, and what it means for engineers

Executive overview

Most AI coding tools assist with text completion. Devin is a fully autonomous agent: it takes a task, spins up its own environment, writes and tests code, and opens a pull request — all asynchronously, like a remote junior engineer.

Cognition's 15-person team runs up to five Devons per engineer simultaneously. Devin already merges roughly 25% of their production PRs; they expect that to exceed 50% by end of year.

The core insight: the bottleneck in software engineering is no longer writing code — it's defining exactly what to build, and that human role only grows more valuable as AI handles execution.

What Devin is and how it works

• Operates fully asynchronously — assign tasks via Slack, Linear, or GitHub issues; Devin opens PRs in your repo
• Builds an internal wiki of your codebase as it works, accumulating knowledge across sessions
• Confidence levels surfaced per task so engineers know when to steer
• Interactive planning phase lets engineers scope work together with Devin before execution
• Devin can start other Devin sessions via API, enabling parallel sub-task execution

How Cognition uses Devin internally

• Each engineer runs up to five Devons concurrently — one per planned task
• Devin handles bug fixes, feature additions, test coverage, and documentation autonomously
• Engineers review and course-correct; the hardest 10–20% of decisions stays with the human
• Devin uses its own product to test code changes — it logs into Devin and starts sessions against the Devin codebase
• New engineers are onboarded partly through Devin's wiki, reducing reliance on senior staff

Bricklayer to architect: how the engineer role changes

• Bricklayer work — debugging Kubernetes errors, migrating code, writing boilerplate — is increasingly delegated to Devin
• Architect work — defining the problem, choosing tradeoffs, specifying architecture — becomes the core human contribution
• Engineers who can think precisely about abstractions (networking, garbage collection, data models) will leverage AI more effectively than those who cannot
• Learning to code is still essential: understanding what lies beneath the abstraction is what makes specifications accurate
• Jevons Paradox applies: as programming gets cheaper and faster, total demand for programming grows — more engineers will be hired, not fewer

Product and landscape

• Devin is positioned as the dedicated autonomous-agent product, distinct from copilot-style inline tools (Cursor, Copilot) and general-purpose models
• Stickiness, not a hard moat, is the defensibility thesis: Devin's value compounds as it learns a team's codebase, process, and conventions
• Devin Search and Devin Wiki are standalone tools that complement the core agent experience
• Linear integration allows one-click task delegation: add a "Devin" label, Devin scopes the ticket and starts work
• Best-fit tasks are well-defined with a clear verification path (front-end changes, bug fixes, test additions); open-ended architectural questions require more human steering

Getting adoption inside a company

• Start with early adopters who set up repos, CI, and linting access for Devin
• Give Devin simple, well-scoped tasks first so it can build codebase familiarity
• Once teammates see Devin's PRs shipping, organic adoption follows naturally
• The mental model that works: treat Devin as a new junior engineer, not a search tool or chatbot

Founding and build approach

• Cognition started as a November 2023 hackathon; official company January 2024; launched March 2024; self-serve December 2024
• Early bet: reinforcement learning (high-compute RL, not imitation learning) would unlock the next capability leap — code was the natural domain because it has an automated feedback loop
• Eight internal pivots within "coding agents" before landing on the current product form
• Team of ~27; 18 have previously founded companies; many share competitive programming backgrounds
• Hiring philosophy: fight for the right person with the same intensity you apply to product — including flying to candidates' families to work through concerns