Kuaishou's KAT-Coder-V2.5 trains coding agents on 100,000 verified environments

The KwaiKAT team at Kuaishou released a technical report for KAT-Coder-V2.5, an agentic coding model. Their AutoBuilder tool raised environment-construction success from 16.5% to 57.2%, producing over 100,000 verifiable repository environments across 12 languages, while a sandbox audit cut reinforcement-learning (RL) feedback errors from about 16% to under 2%.

Why it matters: This is evidence that agentic coding progress is increasingly bottlenecked by training infrastructure -- how reliably you can build and verify realistic coding environments -- rather than raw model scale. It reflects an infrastructure-first thesis several labs have been converging on, suggesting the next round of coding-agent gains may come from environment engineering rather than bigger models.