The Agent Writes the Agent Now
Devin writes 89% of Cognition's own code - the threshold where coding agents become self-sustaining just arrived.
Coding agents crossed a threshold this week that we will not uncross. Cognition disclosed that Devin now writes 89% of the company's own codebase. Not customer code. Its own.
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. Itau auto-fixes 70% of security vulnerabilities. These are not pilot numbers.
What the 89% Actually Means
When your coding agent writes nearly all of its own code, the codebase becomes an agent-native artifact. The code is structured for the agent that wrote it, not for human maintainers browsing files in an IDE.
This is different from autocomplete-style AI assistance. Cursor, Copilot, and Claude Code augment a human-driven workflow - the developer remains the author, the agent accelerates keystrokes. Devin inverts that relationship. You assign a ticket. The agent plans, codes, tests, opens a PR, and iterates on review feedback. The human is the reviewer, not the writer.
That split - agent-first versus IDE-augmented - is now the defining architectural divide in developer tooling. Both work. They produce fundamentally different codebases.
What Changes for Builders
If you build with coding agents - and I do, daily - the 89% figure forces a question: are you optimizing your codebase for human readability or agent operability?
Today, most teams still write code that is human-first. Style guides, naming conventions, folder structures - all designed for people scanning files. An agent-native codebase optimizes for different things:
- Deterministic test suites the agent can run as verification gates
- Machine-parseable task definitions instead of ambiguous ticket descriptions
- Commit histories structured for agent context retrieval
The teams shipping fastest will be the ones that accept the codebase is no longer primarily human-authored and design accordingly.
Where This Heads
The 89% is not the ceiling. It is the current floor for a company that has been iterating this model for two years. Within a year, the concept of a "human-authored codebase" at agent-first companies will sound as dated as hand-written assembly. The question is not whether agents will write most production code. It is whether your engineering process is ready for it.