Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 31·5 min read
openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents
TL;DR
openJiuwen is an open-source harness designed for long-horizon coding agents, enabling dynamic task execution and developer composability. It outperforms existing benchmarks, achieving 82.6% and 87.19% on SWE-bench and Terminal-Bench 2.1, respectively.
✦ Why It Matters
Consider integrating openJiuwen into your coding agent projects for improved adaptability and performance.
Key Takeaways
- Explore openJiuwen for building complex coding agents.
- Utilize Structural Composability to streamline agent development.
- Implement Runtime Adaptivity for dynamic task execution.
- Evaluate your agents using SWE-bench and Terminal-Bench metrics.
How It Works
openJiuwen operates by providing a shared execution substrate that allows for Rail-based capability composition. This means developers can create sophisticated agent systems that can dynamically adjust their behavior based on real-time evidence, such as task progress and execution outcomes.
The framework enables agents to make informed decisions during runtime, enhancing their ability to complete complex coding tasks.