One of the most widely read articles on VentureBeat recently has focused on Devin, an AI “software engineer” developed by startup Cognition. Devin has the remarkable ability to autonomously generate code, including entire applications, based on natural language instructions provided by users.
This capability holds significant promise for software developers, managers, and their employers. Influential figures in the AI community, such as University of Pennsylvania Wharton School of Business professor Ethan Mollick, have showcased examples of Devin successfully coding complete websites under their guidance.
However, for enterprises operating in highly secure and regulated sectors like finance, healthcare, law, government, and telecommunications, the prospect of unleashing an AI software engineer independently to handle tasks related to maintaining existing codebases and applications, as well as implementing updates for employees or customers, may raise concerns.
In response to this, an emerging Israeli startup called Codium AI is striving to offer enterprises the benefits of autonomous AI coding while ensuring human oversight and compliance with regulations and existing software standards.
Recently, Codium AI introduced “Codiumate,” a new AI agent described as “semi-autonomous.” This tool collaborates with human software developers or teams by analyzing highlighted code snippets in their environment. It then automatically generates step-by-step development plans, writes code based on the plan, identifies duplicate code segments, creates code documentation for future reference, and proposes tests to validate the code before deployment in live production environments.
According to Codium’s co-founder and CEO Itamar Friedman, developers spend only about “20 to 30 percent of their time” on actual coding tasks, as indicated by surveys like the one conducted by The New Stack and Tidelift. The rest of their time is dedicated to defining the code’s purpose, verifying its implementation, and ensuring its functionality—a focus that Codium AI aims to support.
The Codiumate tool was demonstrated live by Friedman during a video call, illustrating how users can select multiple code blocks in their development environment using their mouse cursor and instruct Codiumate to generate a corresponding plan. The generated plan is then reviewed by the human developer, who can modify it as needed. Upon approval, the human developer can initiate the code implementation process with Codiumate’s assistance, allowing for real-time adjustments and interventions as required.
Codiumate leverages advanced technologies, including “best of breed” OpenAI models, Anthropic’s Claude, and Google’s Gemini. Additionally, Codium has developed its own large language model (LLM) through the AlphaCodium process, enhancing the performance of other LLMs in code completion tasks. While the general models are accessible to all users, the Codium LLM is exclusively available to enterprise clients, offering superior coding capabilities compared to OpenAI models.
To address data security concerns, Codiumate follows a “zero retention” policy, ensuring that all code and information entered into the tool are promptly deleted once the user ends their session. The company adheres to industry standards such as SOC 2 and provides the option for running Codium on private servers or air-gapped computers for enhanced data protection.
In terms of pricing and availability, Codiumate is part of Codium AI’s expanding suite of AI development agents. The offerings include a free tier powered by OpenAI’s GPT-3.5, a Teams tier priced at $19 per user per month, and an Enterprise tier with customizable software-as-a-service (SaaS) subscription plans tailored to individual customer requirements. Codium AI boasts a user base of 500,000 globally, as per a company press release.
