KAT-Coder-Pro V2.5 Launch: Boosting Agentic Coding from Code Writing to Full Engineering

Since the first release of KAT‑Coder‑Pro V2, developers have asked whether a model can take a full issue, locate the relevant files in a large repository, apply changes, run tests and handle failures autonomously. Existing coding models typically excel at atomic tasks such as fixing a single function, but fall short on the end‑to‑end engineering workflow.

Long‑Chain Engineering Capability

The new model, KAT‑Coder‑Pro V2.5, breaks three key bottlenecks: (1) cross‑file location errors, (2) failure to follow project conventions, and (3) inability to run and verify tests. Building realistic project environments is hard; automatic construction succeeds only about 16.5% of the time.

To address this, the team built AutoBuilder , an automated pipeline that lets the model act as an “environment‑building engineer”. AutoBuilder repeatedly analyses repository structure, generates configuration scripts, validates tests in an isolated sandbox, and iterates to fix failures. This pipeline raises environment‑construction throughput by 57.2%, covers 12 programming languages, and creates over 100 000 runnable, verifiable repository environments spanning defect fixing, feature addition, API compatibility, cross‑module integration and regression fixing.

On the data side, a full‑process behavior filter identifies high‑value failure trajectories—cases where the model located the right files and direction but missed a critical decision. About 20% of these are converted into complete, reproducible training samples, teaching the model not only how to succeed but also how to recover from mistakes.

General Agentic Framework

Beyond long‑chain engineering, the KwaiClawEnv system provides a universal Agentic training environment. It consists of three layers:

Service layer : dynamically expands the tool pool by converting open‑source skill definitions into deployable services, avoiding a fixed‑tool limitation.

Task layer : seeds real business tasks and generates massive task variants through tool‑call chains, dependency constraints and difficulty parameters.

Eval layer : applies hard rules plus model‑based review to retain only executable, verifiable, and behaviorally natural trajectories.

The resulting training data span data analysis, cross‑system integration, batch document processing and report generation—core daily needs of thousands of Kwai developers.

Large‑Scale Agentic Reinforcement Learning

Supervised fine‑tuning (SFT) can teach a model to copy seen trajectories, but it fails on unseen project structures, unexpected errors, or tool feedback. To endow true exploration, error‑correction and verification abilities, the team applied large‑scale RL with three innovations:

Harness Scaling : diversified scaffolds across multiple Agent frameworks (mini‑swe‑agent, Claude Code, Codex, OpenClaw) to prevent over‑fitting to a single tool protocol.

Long‑Chain Credit Assignment : an asymmetric PPO architecture where the online actor sees only the deployed environment, while the critic accesses “god‑view” signals such as final test pass/fail and patch quality, enabling precise reward attribution for each step.

Hierarchical Reward System : three‑tier rewards—core task score based on real test outcomes, standard behavior constraints to penalise malformed tool calls or leftover artifacts, and failure‑trajectory incentives that reward partial successes like correct file localisation or partial test passes.

Multi‑Expert Fusion (MOPD)

Five specialist models were trained for long‑chain repository engineering, general Agentic tasks, end‑user interaction, front‑end aesthetic generation, and general knowledge. Using Multi‑Teach On‑Policy Distillation (MOPD), a single student model learns from the appropriate expert for each question, merging capabilities in function space rather than raw parameters. The resulting V2.5 model retains all prior abilities without the “see‑saw” effect of sacrificing one skill for another.

Benchmark Results

On SWE‑Bench Pro, KAT‑Coder‑Pro V2.5 scores 65.2; on the internal KAT Code Bench it reaches 53.1. In multi‑round Agentic evaluation (PinchBench) it attains 94.2, and on the internal Agentic benchmark (KAT Claw Bench) it scores 85.5, demonstrating stable performance across tool selection, context management and final deliverable generation.

The model is now fully released via the StreamLake platform; developers can call the API directly, consult the technical report (arXiv:2607.05471) for full details, and join the KwaiKAT community for updates.