Three weeks after the primary software (SW) development ‘artificial intelligence (AI) agent’ appeared, a second AI agent appeared.
Enterprise Beat reported on the third (local time) that Codium, a code generation AI startup, unveiled ‘Codium Mate.’
In keeping with this, Cordium Mate is an AI agent for SW development just like ‘Devin’ introduced by Cognition AI three weeks ago.
Devin is fully autonomous, able to handling the whole development project by itself, from writing code to fixing bugs to running the ultimate product, with none human assistance. Nevertheless, Codium Mate is a ‘semi-autonomous’ agent that works with human developers.
In other words, the developer first provides Codeumate with existing code fragments that will likely be used as a reference for the SW to be developed. Based on this, Codium Mate drafts a step-by-step development plan, generates code based on the plan, identifies duplicate code, and creates a draft document for the code for developers to discuss with. Finally, testing is finished to be sure that the code works properly before it is definitely deployed to the actual production environment.
The important thing on this process is to report back to humans at vital decision-making points and follow their decisions. In reality, Codium Mate presents a plan to developers at each project stage, and humans can freely modify and adjust it.
That is effective when there may be a necessity to keep up existing codebases and apps, or when appropriate human intervention is required in highly secure and highly regulated fields equivalent to finance, healthcare, law, government, and telecommunications.
In other words, the developer acts as a manager who instructs, supervises, and checks the AI agent, and the AI agent replaces the present mechanical and time-consuming role.
Itama Friedman, co-founder and CEO of Cordium, said, “In real projects, writing code is simply 20 to 30 percent of a developer's work, and most of it’s defining what you do with the code or what you would like to do with the code and recording what actually happened. “We focused on this and built Cordium Mate semi-automatically,” he explained.
The technology applied to Cordium Mate is OpenAI's 'GPT-4 Turbo', and Antropic's 'Claude 3' and Google's 'Gemini' are also being tested.
To enhance performance, the self-developed Large Language Model (LLM) can be used, which has superior coding performance than existing models and is provided to paid corporate users. He added that a Fortune 10 company is already using it.
Meanwhile, Codium developed an AI-based coding helper tool last January and succeeded in attracting investment price $65 million (roughly KRW 88 billion) with a company value of $500 million (roughly KRW 680 billion), resulting in the code generation AI boom.
Codium's AI coding toolkit built LLM into the integrated development environment (IDE) and gained great popularity as a tool that mechanically generates code by inputting natural language descriptions or supports translation between programming languages. It is alleged that there are currently 500,000 users.
However, Cognition created a worldwide sensation with Devin's disclosure, and its corporate value rose to $2 billion (roughly 2.7 trillion won) in two weeks, and is within the strategy of making recent investments.
As AI agents appear one after one other in the sector of code generation, expectations are growing. Code generation is an area where open AI, Microsoft, Google, and Amazon are fiercely competing. They’ve also stated that their goal is to develop AI agents.
Reporter Lim Da-jun ydj@aitimes.com