We present KAT-Coder-V2.5, a coding-focused agentic model designed to operate autonomously within real, executable repositories rather than merely serving as a single-turn code generator. Its capabilities are constrained by the scarcity of reproducible environments, verifiable rewards, and high-value trajectories, which we address through an end-to-end agentic post-training framework.
AutoBuilder reconstructs multilingual repositories into sandboxed environments with fail-to-pass and pass-to-pass verification at scale, from which we regenerate self-contained task specifications, recover near-miss trajectories, and distill supervision through process-aware filtering. Meanwhile, KwaiClawEnv synthesizes large-scale tool-use trajectories from executable services and real task seeds.
We further scale reinforcement learning using harness randomization, a reliability-hardened sandbox, an asymmetric actor-critic PPO with hindsight-augmented value estimation, and a harness-oriented reward framework, unifying SWE, Agent-Claw, and WebCoding experts via Multi-Teacher On-Policy Distillation. Across six software engineering and agentic benchmarks, KAT-Coder-V2.5 delivers the best agentic tool-use results on PinchBench and ranks second only to the frontier Opus 4.8 in repository-level software engineering.
Our service is available at KAT-Coder.
Blogger's Review: KAT-Coder-V2.5 represents a significant breakthrough in autonomous coding capabilities, addressing critical pain points in model training by simulating real-world environments, thus providing new insights for the development of future coding tools.