Purpose Devin is the "world's first fully autonomous AI software engineer" launched by Cognition, positioned as a tireless, highly skilled team member who can both build alongside engineers and independently complete tasks for users to review. According to the official description, with Devin, engineers can focus on more interesting problems, and engineering teams can pursue larger goals.
Main capabilities
- Long-term reasoning and planning: According to the official description, with advances in long-term reasoning and planning, Devin can plan and execute complex engineering tasks that require thousands of decisions, and can recall relevant context at each step, learn over time, and correct errors.
- Developer tools and sandboxed environment: Devin is equipped with common developer tools, including a shell, code editor, and browser, and runs in a sandboxed computing environment, which the official description calls "everything a human needs to do work."
- Real-time collaboration: Devin can report progress in real time, accept feedback, and work with users on design choices as needed.
Examples of capabilities given officially
- Learning unfamiliar technologies: After reading a blog post, Devin ran ControlNet on Modal to generate an image for Sara with a hidden message.
- Building and deploying an application end to end: Devin created an interactive website simulating the "Game of Life," gradually added features requested by the user, and then deployed the application to Netlify.
- Autonomously finding and fixing bugs: Devin helped Andrew maintain and debug his open source competitive programming book.
- Training and fine-tuning its own AI model: Based only on a link to a research repository on GitHub, Devin set up fine-tuning for a large language model.
- Handling issues and feature requests in open source repositories: Given only a GitHub issue link, Devin completed all necessary setup and context gathering.
- Contributing to a mature production repository: In one example from the SWE-bench benchmark, Devin fixed a logarithmic calculation bug in the sympy Python algebra system - it set up the code environment itself, reproduced the bug, and wrote and tested a fix.
- Taking on real work: The official description says they tried giving Devin real Upwork work, and it could also complete it; in one example, Devin wrote and debugged code to run a computer vision model, sampled the resulting data, and compiled a report at the end.
Performance The official description evaluated Devin on SWE-bench. This benchmark requires the agent to solve Django, scReal GitHub issues in open-source projects such as ikit-learn. Officially, Devin correctly resolved 13.86%* of the issues end-to-end, far exceeding the previous best of 1.96%; even when told the exact files that needed to be edited, the previous best model could only resolve 4.80% of the issues. Officially noted, Devin was evaluated on a random 25% subset of the dataset, Devin was evaluated without assistance, while all other models were with assistance (that is, told exactly which files needed to be edited). Officially plans to release a more detailed technical report soon.
About Cognition Cognition is an applied AI lab focused on reasoning and is building AI teammates whose capabilities far exceed those of today's existing AI tools. Officially, by solving reasoning problems, new possibilities can be unlocked across a wide range of disciplines—code is only the beginning. Cognition hopes to help people around the world turn ideas into reality. The company is well funded, including a $21 million Series A round led by Founders Fund, and thanks industry leaders including Patrick and John Collison, Elad Gil, Sarah Guo, Chris Re, Eric Glyman, Karim Atiyeh, Erik Bernhardsson, Tony Xu, Fred Ehrsam, and others for their support.
Usage and limitations Officially, Devin is currently in early access, and the company is increasing capacity. To use Devin to complete engineering work, you can join the waitlist or contact info@cognition.ai. Regarding specific fees, pricing, platform support, and detailed usage steps, the official introduction does not currently provide them.
