Gemini Code Assist, Google’s AI coding assistant, is unveiling new “agentic” capabilities in preview.

During its Cloud Next conference on Wednesday, Google announced that Code Assist can now deploy new AI “agents” that are able to take multiple steps to tackle complex programming tasks. These agents can develop applications from product specifications in Google Docs, or even perform code transformations from one language to another. Code Assist is now accessible in Android Studio, as well as other coding environments.

The upgrades to Code Assist are likely a response to the competitive pressure from rivals like GitHub Copilot, Cursor, and Cognition Labs, the creators of the viral programming tool Devin. The AI coding assistant market is becoming increasingly competitive, and there’s a significant amount of money involved. Given the potential for boosting productivity, this doesn’t come as a surprise.

Code Assist’s agents, which can now be managed from a new Gemini Code Assist Kanban board, can generate work plans and provide step-by-step updates on job requests. In addition to creating software and migrating code, the agents are capable of implementing new app features, conducting code reviews, and even generating unit tests and documentation, as claimed by the company.

However, the effectiveness of Code Assist in performing these tasks remains uncertain. Studies have shown that even the most advanced code-generating AI often introduces security vulnerabilities and bugs due to weaknesses in understanding programming logic. A recent evaluation of Devin found that it successfully completed only three out of 20 tasks.

Therefore, if you decide to use Code Assist to create or refactor an app, it might be a good idea to review the code yourself for added safety.