Devin 2.2 and Beyond: Cognition’s Latest AI Agent Enhancements

Cognition has been steadily rolling out updates to Devin, its cloud-based agent platform for engineering teams. The latest release, Devin 2.2, introduces significant improvements in task scheduling, parallel processing, and state management. These changes are designed to make Devin more versatile and efficient, allowing it to handle complex tasks with greater autonomy. Let’s dive into the technical details and see what these updates mean for practitioners.

Devin Can Now Schedule Recurring Sessions

One of the key features introduced in the recent update is the ability for Devin to schedule its own recurring sessions. This means that after a task has been successfully executed, you can instruct Devin to continue running it on a regular basis. Here are the highlights:

• State Maintenance: Each session picks up where the previous one left off, ensuring continuity and reducing redundant work.
• Automation of Routine Tasks: Ideal for tasks like data synchronization, report generation, or periodic code reviews.

This feature is particularly useful for teams that need to perform repetitive but critical operations without manual intervention. By maintaining state between runs, Devin ensures that each session builds on the progress of the last, making it a powerful tool for long-term projects.

Managing Multiple Devins in Parallel

Another significant enhancement is Devin’s ability to manage multiple instances of itself, each running in its own isolated virtual machine (VM). This parallel processing capability allows for:

• Task Decomposition: Large tasks can be broken down into smaller, manageable sub-tasks.
• Resource Isolation: Each Devin runs in a separate VM, ensuring that resource usage is controlled and predictable.
• Concurrency: Multiple Devins can work on different parts of the same task simultaneously, significantly reducing overall execution time.

This feature is particularly beneficial for complex projects where tasks can be parallelized. By running multiple Devins in isolation, teams can ensure that each sub-task is handled efficiently without resource contention or interference.

Early Preview of SWE-1.6 Training

Cognition has also shared an early preview of its ongoing SWE-1.6 training run. This update focuses on improving the performance and capabilities of Devin’s underlying AI models. Key points include:

• Enhanced Model Training: The team is working on refining the model to better understand and execute complex engineering tasks.
• Performance Benchmarks: Early results show promising improvements in task completion time and accuracy.

While still in progress, this training run aims to make Devin even more effective at handling a wide range of engineering challenges.

How Cognition Uses Devin to Build Devin

In a fascinating blog post, the Cognition team shared how they use Devin to build and improve itself. This meta-usage highlights:

• Task Delegation: Engineers can delegate tasks like code reviews, bug fixes, and documentation updates to Devin.
• Collaborative Development: Devin works alongside human engineers, contributing to PRs and handling backlog items.

This approach not only accelerates development but also ensures that the team can focus on high-priority tasks while Devin handles routine work.

Cognition for Government

Cognition has launched a specialized version of its platform, Cognition for Government, aimed at modernizing critical infrastructure with AI software engineering. This initiative is designed to:

• Enhance Security: Ensure that government systems are built and maintained securely.
• Improve Efficiency: Streamline development processes to reduce costs and accelerate project timelines.

This move underscores Cognition’s commitment to applying advanced AI technologies in public sector projects.

Closing the Agent Loop: Devin Autofixes Review Comments

A recent feature addition allows Devin to automatically fix review comments, which has significantly improved code quality. Key benefits include:

• Increased Token Spend: While this feature has led to a higher internal token spend on Devin, it has also resulted in cleaner and more bug-free PRs.
• Enhanced Productivity: Engineers can now focus on more strategic tasks, knowing that Devin is handling the nitty-gritty of code reviews.

This feature demonstrates Cognition’s ongoing commitment to enhancing the practical utility of AI agents in real-world development environments.

Agent Trace: Capturing the Context Graph of Code

Cognition has joined forces with Cursor, Cloudflare, Vercel, git-ai, and OpenCode to support Agent Trace, an open spec for recording AI contributions alongside human authorship in version-controlled codebases. Key features include:

• **