Krafton开发内部AI招聘评估工具Cofa-Probe，用于评估开发者与AI协作能力

Krafton (259960), which has made "AI first" a company-wide principle, has developed its own AI-based hiring and assessment solution. The aim is to judge candidates not only on what they produce with AI but also on how they work with it, using objective measures to identify the right talent.
Krafton has begun using Cofa-Probe, a hiring and assessment solution built by its AI Frontier Division, in recruiting for developer positions, according to information technology industry sources on the 29th. Cofa-Probe is a large language model-based solution that evaluates how candidates define problems and verify results while working with AI. It was developed using OpenAI's Codex.
Cofa-Probe creates a virtual environment that resembles the actual work of an AI engineer. Within that environment, it assesses how candidates identify and solve problems while interacting with multiple task agents. Unlike conventional assessments that look only at the finished product, it structures the entire way a candidate works with AI.
The solution breaks the collaboration process into four categories: problem definition and prioritization, quality of AI collaboration and instructions, evidence-gathering and decision-making, and iterative improvement and error recovery. At each stage, it evaluates whether candidates single out the core problem, clearly convey roles, context and expected outcomes to the AI, build hypotheses based on data and evidence, and trace the causes of failure to see results through to the end.
During assessment, Cofa-Probe automatically collects not only source code but also conversation logs, prompts, tool settings and traces of verification runs. After a task ends, candidates take part in a follow-up question-and-answer session explaining their output and decision-making process, and the system generates a reference report for interviewers.
The company also built in safeguards to address the limits of LLM-based grading. With an LLM, the same submission can yield different results from one run to the next. To prevent that and improve the reproducibility and discriminating power of the assessment, Krafton applied a "meta harness" that repeatedly tests and refines the execution structure of the AI grading agent. Elements that require consistency, such as evaluation criteria and score conversion, are controlled by code, while qualitative judgments that require reading context are handled by LLM-based agents.
Krafton developed Cofa-Probe to properly identify the AI-native talent it wants. The company declared in October last year that it would become an AI-first firm and has been devoting company-wide resources to its AI transformation. "Cofa-Probe began with the question of how we can verify the capabilities that talent needs in the AI-native era," a Krafton official said. "We wanted to turn the process of collaboration between humans and AI into an observable structure so that we could assess problem-solving ability that does not show up in the output alone."
Krafton plans to expand the scope of Cofa-Probe. To gauge that potential, the company held Cofathon, a hiring-linked AI-native hackathon, with CJ Olive Young on the 30th of last month at PUBG Seongsu in Seoul's Seongdong district. Participants carried out one of two practical tasks presented by Krafton and Olive Young. Krafton applied the Cofa-Probe system in grading the tasks, presenting observable evidence of problem-solving approaches and AI proficiency that are difficult to gauge from the final product alone.
"Competitiveness in the AI era comes from defining problems together with AI, checking the evidence and verifying results all the way through," said Park Jae-min, head of Krafton's AI Frontier Division. "We will advance Cofa-Probe to set a new standard for assessing AI-native talent."