cd ~/
Agentic AIAI EngineeringDeveloper ToolsAgent ArchitectureAutonomous Software

Devin Writes 89 Percent of Cognition's Code

Devin writes 89% of Cognition's own code - the threshold where coding agents become self-sustaining just arrived.

Evyatar Bluzer
3 min read

Devin now writes 89% of Cognition's own codebase. The company disclosed that number this week, and it marks the kind of threshold you do not get to uncross: the agent vendor's own code, written by the agent, at production scale.

The Numbers

Cognition raised $1 billion at a $26 billion valuation on May 27. The round is large but the operating numbers are what matter:

  • $492 million annualized revenue, up from $37 million twelve months ago - a 13x jump
  • 89% of committed code at Cognition is written by Devin, up from 13% in December 2025
  • 50% month-over-month growth in enterprise usage, sustained for six months
  • Customers include Mercedes-Benz, Goldman Sachs, NASA, the US Army

Mercedes-Benz compressed an eight-month legacy modernization into eight days, and Itau auto-fixes 70% of security vulnerabilities. These are not pilot numbers.

What Does 89% Actually Mean?

When your coding agent writes nearly all of its own code, the codebase becomes an agent-native artifact, structured for the agent that wrote it rather than for human maintainers browsing files in an IDE.

That differs from autocomplete-style AI assistance. Cursor, Copilot, and Claude Code augment a human-driven workflow, where the developer remains the author and the agent accelerates keystrokes. Devin inverts the relationship: you assign a ticket, the agent plans, codes, tests, opens a PR, and iterates on review feedback, and the human's job is review. That split - agent-first versus IDE-augmented - is now the defining architectural divide in developer tooling. Both work. They produce very different codebases.

Human-Readable or Agent-Operable?

If you build with coding agents - and I do, daily - the 89% figure forces you to pick which of those two properties your codebase optimizes for. Most teams still write human-first code: style guides, naming conventions, folder structures, all designed for people scanning files. An agent-native codebase optimizes for different things:

Agent-first versus IDE-augmentedA highlighted vertical line divides two columns of text; each column has a heading, a short description of who authors and who reviews, two or three optimized-for items, and a footer line. IDE-AUGMENTED, HUMAN-FIRST CODEBASE AGENT-FIRST, AGENT-NATIVE CODEBASE the defining divide Cursor, Copilot, Claude Code the developer is the author, the agent accelerates keystrokes Devin the agent plans, codes, tests, opens the PR, and iterates on review; the human reviews OPTIMIZED FOR OPTIMIZED FOR Style guides and naming conventions Folder structures for people scanning files Deterministic test suites the agent runs as verification gates Machine-parseable task definitions instead of ambiguous ticket descriptions Commit histories structured for agent context retrieval where most teams still write code Cognition: 89% of committed code written by Devin, up from 13% in December 2025
The two codebases: on the left the developer authors and the agent accelerates, on the right Devin authors and the human reviews, and each side is optimized for a different reader.

Teams that accept the codebase is no longer primarily human-authored, and design for that, will ship faster than the teams still pretending otherwise.

The Floor

Treat 89% as a floor rather than a ceiling; it is where a company lands after two years of iterating on this model. Within a year, a "human-authored codebase" at agent-first companies will sound as dated as hand-written assembly. Whether agents will write most production code stopped being an open question a while ago. Whether your engineering process is ready for it is the one that remains.

Comments