Cognition

SWE-2

SWE-2 is an advanced agentic coding model designed to push the Pareto frontier of capability and cost-efficiency. It utilizes a novel reinforcement learning (RL) algorithm that trains across multiple reasoning-effort levels in a single run, enabling high-performance software engineering tasks at a fraction of the cost of previous frontier models.

Agent Memory Repo

Agent Memory Repo is an open specification that enables AI agents to maintain persistent, cross-session memory by utilizing a Git repository as the storage backend. It organizes information through a structured layout of Markdown files, allowing agents to read, update, and link memories like a wiki.

SWE-1.7

SWE-1.7 is a highly capable AI model developed by Cognition, optimized for long-horizon asynchronous software engineering tasks. Built on a Kimi K2.7 base, it leverages advanced reinforcement learning techniques to achieve frontier-level intelligence while significantly improving cost-efficiency.

Devin Fusion

Devin Fusion is a multi-model harness developed by Cognition that integrates frontier-level AI models with cost-effective 'sidekick' agents, achieving 35% lower costs while maintaining high performance in software development tasks.

Devin Desktop

Devin Desktop is an integrated development environment (IDE) designed to manage both local and cloud-based AI agents seamlessly. It allows developers to plan, delegate, review, and deploy code without leaving their editor, enhancing workflow efficiency and collaboration.

Devin 2.2

Cognition has officially launched Devin 2.2, a major update to its autonomous AI software engineer that significantly expands its capabilities for end-to-end project management. This release introduces advanced computer operation features, allowing Devin to perform real-world testing, self-verification, and automatic error correction directly within its environment. In addition to a three-fold increase in startup speed and a complete user interface overhaul, Devin 2.2 deepens enterprise integration with enhanced support for Slack and Linear and introduces an integrated self-review system to ensure higher code quality. By closing the "agent loop" with these autonomous verification tools, Cognition continues to push the boundaries of AI-driven engineering, enabling more reliable and complex software development workflows.