The year 2025 will be remembered in future technology history as "the year software engineering was redefined."
As CTO of Japan AI, when I look back on this turbulent year, what comes to mind isn't feature lists or KPI numbers. It's the shock of DeepSeek benchmarks shared by engineers in late-night Slack, the "spine-chilling sensation" when first seeing Claude 3.5's output code, and the pain of burying parts of products we had built with our own hands.
This is the record of a technology organization that decided to walk alongside AI, this "new species."
The changes that hit the industry in early 2025 weren't gentle "evolution"—they were "disruption."
What particularly stands out in my memory is when Claude 3.5 (and later 3.7) and Claude Code emerged. Until then, I had been somewhat skeptical of AI coding tools. "In the end, humans will have to fix it, right?" But that thinking was shattered overnight. The code AI generated didn't just work—it had more elegant structure than that of inexperienced engineers.
In that moment, with a mix of fear and excitement, I made one decision: "Contract Claude Code MAX Plan for all engineers." This wasn't just tool adoption—it was a declaration to the team that "we're not going back to the old way of writing."
Then came DeepSeek. This was a historic benchmark for the industry: "reasoning ability becoming commoditized." Even open-weight models became capable of "reasoning," and AI was no longer a probabilistic parrot. They worry, correct, and derive answers.
(Note: In our production systems, we strictly manage model selection and deployment environments with consideration for safety and data sovereignty.)
Seeing this change, I arrived at one hypothesis: "Intellectual tasks with clear goals and evaluation criteria become prime candidates for Agent automation."