DeepSWE v1.1

DeepSWE v1.1 is an AI model that independently repairs and develops software further. It is given a task in text form – for example, the description of a bug in a program – and the model itself searches for the cause, writes new code, and checks the result. It does not act step by step according to fixed instructions, but rather plans its own course of action. This places it in the class of so-called AI agents: systems that do not merely answer a task, but actively work on it across multiple steps. It was developed by Cognition AI, an American start-up specializing in such autonomously acting programming systems.

DeepSWE v1.1 and the race for autonomous programming

The benchmark in this field is SWE-bench Verified, a standardized test for autonomous programming systems. It consists of real, reported bugs from public open-source projects on GitHub – in other words, real problems that real developers could not simply solve. A model passes the test if it fixes the bug completely and correctly, without outside help.

DeepSWE v1.1 solves around 60 percent of all tasks there. This figure sounds unspectacular, but it is extraordinary by comparison: as recently as early 2024, 20 percent was considered a good result. DeepSWE v1.1 is among the first publicly accessible models to far exceed this mark – demonstrating just how fast the field is moving. Its significance thus lies not only in the product itself, but in the pace of development.

How DeepSWE v1.1 conducts a bug search

The foundation is a pretrained language model – a system that understands and generates text, similar to a chatbot. On top of this, a layer was added that prompts the model to proceed like a software developer: reading files, making targeted changes to code, running tests, evaluating error messages, and trying a different approach if needed.

This behavior was trained using reinforcement learning, a learning method in which the model tries out tasks and receives a reward if the solution is correct – and none if it is not. It therefore does not learn from pre-made sample solutions, but through trial and error. This is unusual for programming systems, because it requires no intermediate evaluations: either the test passes, or it does not.

An important difference from ordinary code assistants: DeepSWE v1.1 works on an entire project, not on a single code snippet. It navigates through folder structures, selectively reads relevant sections, and decides for itself which files are even relevant. This brings it closer to the approach of a human developer than earlier systems.

DeepSWE v1.1 in practice and in the headlines

In software development, such systems are predicted to play a major economic role. The idea is not that they will replace human developers, but that they will take over routine tasks: fixing small bugs, updating outdated libraries, automatically writing tests. This gives developers time for more complex problems.

In technology news, DeepSWE v1.1 appears as an example of how quickly the performance of such agents is increasing. Investors and analysts now watch the SWE-bench score in a manner similar to how language-understanding benchmarks were once tracked – as a metric for how close AI systems are coming to real work. Cognition AI itself operates a product called Devin based on such models, marketed as the first commercial “AI software engineer”.

Critics point out that SWE-bench only captures a fraction of reality. Real projects are larger, messier, and often contain unwritten rules that no test captures. DeepSWE v1.1 is a milestone – but not proof that autonomous programming already works reliably at large scale.