Keep pulling the thread on Scott Wu.
Approximately 25% of pull requests at Cognition are currently created by the Devon AI agent.
Cognition expects the Devon AI agent to be responsible for more than half of its pull requests by the end of the year.
The integration between Devon and Linear allows users to trigger the AI agent to work on a task simply by adding a "Devon" label to a ticket in Linear.
Scott Wu predicts that within several years, the software development experience with agents like Devon will evolve to a point where users can modify applications by giving natural language commands directly on the product interface, without looking at the code.
Scott Wu believes that current foundation models have sufficient "base intelligence" for AI engineering agents, and the primary challenge is now teaching them the specific, idiosyncratic details of real-world engineering workflows.
Cognition's engineering team consists of approximately 15 people.
Most engineers at Cognition work with up to five instances of the Devon AI agent simultaneously.
The AI agent Devon merges several hundred pull requests into production for Cognition's own codebases each month.
Cognition pivoted approximately eight times within the domain of coding agents before settling on the current product direction for Devon.
The Devon AI agent is designed to integrate into existing engineering workflows, allowing users to assign it tasks via Slack and Linear, and it will submit pull requests on GitHub.
The capability of the Devon AI agent has progressed from the equivalent of a high school CS student at launch, to a college intern, and is now at the level of a junior engineer.
Cognition's customer base for Devon ranges from one-to-two person startups to Fortune 100 companies and public banks.