Keep pulling the thread on Jared Zoneraich.
More Devin agents are now launched programmatically or by other Devin agents than by human users.
Devin is particularly effective for "brownfield" tasks such as large-scale code migrations, increasing test coverage, and other forms of developer "grunt work."
In Devin's architecture, each child agent runs in its own dedicated virtual machine, complete with its own computer, display, and ability to test its own code.
Cognition has developed its own proprietary model called Sweet, which is specifically designed for software engineering tasks to be faster and cheaper.
Jared Zoneraich predicts that code will become a "second-order artifact" as developers begin interacting with higher-level abstractions managed by AI agents.
Cognition's strategic view is not to replace coding but to create an abundance of good software by multiplying the effectiveness of engineers.
Jared Zoneraich believes that in the future, "tool engineering," or determining which tools to provide to an AI model, will become a key discipline.
Cognition's defensibility is built on its agent harness and its go-to-market motion involving Forward Deployed Engineers (FDEs).
The Devin platform is architected as a cloud-native, asynchronous system from day one.
Devin automatically routes tasks to the best available model, removing the need for users to select a model manually in the cloud interface.
The most common way users interact with Devin is through Slack and other automations.
Jared Zoneraich predicts that reliance on complex prompting hacks for AI agents will decrease as models improve, making such hacks a poor source of long-term defensibility.