Code Mate user base continues to expand

Monthly token usage rose 3.5 times from May to last month

Developed in 2023, runs on internal infrastructure for security

Search capabilities upgraded to surface past code and technical documents

Covers code analysis, revision and result verification

Samsung Electronics' Hwaseong campus in Hwaseong, Gyeonggi Province. Photo by Lim Se-jun
Samsung Electronics' Hwaseong campus in Hwaseong, Gyeonggi Province. Photo by Lim Se-jun

Samsung Electronics said the user base for its in-house AI coding agent, Code Mate, is growing. Developed in 2023, Code Mate has expanded beyond simply recommending or generating code to analyzing and modifying existing code and running tests.

Samsung Electronics said Thursday that the number of in-house developers using Code Mate has surpassed 10,000. Monthly token usage reached approximately 1.45 trillion tokens last month — about 3.5 times the level recorded in May, when adoption of the AI coding agent began in earnest this year. Tokens are the units AI uses to process text, code and other data.

Code Mate, applied to work since 2023, is a proprietary AI coding agent built with semiconductor development environments and security requirements in mind. It was created to improve development productivity in semiconductor software.

Semiconductor equipment and process software development requires code and specialized knowledge accumulated over many years. General-purpose AI models alone struggle to fully grasp a company's proprietary coding conventions and existing implementation practices.

Samsung Electronics built Code Mate on an on-premises model, running the AI model and service on internal infrastructure. The setup allows developers to draw on internal code and technical documents without any development data leaving the company.

Samsung Electronics' Code Mate on-premises architecture. [Provided by Samsung Electronics]
Samsung Electronics' Code Mate on-premises architecture. [Provided by Samsung Electronics]

Samsung Electronics has also upgraded the information retrieval technology used to find past implementation cases and technical documents during semiconductor software development. Beyond simple keyword matching, the company has researched tools including Semantic Code Finder, which searches for code with similar meaning, and Find My Code Twin, which uses a large language model to identify similar implementation patterns.

Model Context Protocol, or MCP, connects the code and technical knowledge retrieved this way to actual development tasks. MCP is a technology that helps AI access data and tools in a standardized way. When a developer asks Code Mate to solve a problem, the agent uses connected search tools to locate technical documents and similar implementation cases, then analyzes and modifies the code before verifying results through development tools and tests.

Developers have reported that Code Mate has meaningfully cut actual development time. The company said users have noted that tasks previously taking three to four days can now be completed in a single day.

To reduce response delays as the user base grows, Samsung Electronics is expanding GPU resources and optimizing the AI model serving environment. The company is also analyzing the causes of repeated searches and interruptions that occur mid-task and incorporating those findings into performance improvements.

Samsung Electronics plans to develop Code Mate into an AI coding agent capable of handling more complex and extended development tasks. To that end, the company is working to improve AI model performance while strengthening the agent's ability to accurately navigate development context, use appropriate tools and carry out multi-step tasks in sequence.

The company also plans to build an evaluation environment based on real development work to continuously measure and improve Code Mate's task performance. The broader strategy is to strengthen the competitiveness of Samsung Electronics' semiconductor products by boosting software development productivity.


jeongwan@heraldcorp.com