Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview (github.com)

Article summary

The article presents Dirac, an open-source coding agent that topped the TerminalBench-2 leaderboard with a 65.2% score using the gemini-3-flash-preview model. Dirac is designed to optimize for efficiency and cost while maintaining high accuracy, and it achieves this through features such as hash-anchored edits, AST-native precision, and multi-file batching. The agent is built on top of the Cline project and supports various models, including Qwen and other popular open and closed models. Dirac's performance is attributed to its ability to reduce API costs by 64.8% on average while producing better and faster work.

Main themes

• Coding agents
• Efficient coding
• API cost reduction
• AST-based editing
• Model optimization
• Open-source development

What commenters say

• The use of ASTs in coding agents can significantly improve their performance and efficiency, especially in large codebases.
• Some commenters believe that the efficiency of Dirac comes mainly from showing file skeletons by default, rather than its anchor-based editing approach.
• There is a need for more intentional prompting to enable the use of tools like ast-grep and gritql in coding agents.
• The choice of tools and APIs in coding agents should be minimal and opinionated to avoid overloading the model with too many options.
• The performance of coding agents can be improved by incorporating more real-world engineering experience and lessons learned from other industries.
• Some argue that the focus on model training and serving infrastructure is more important than tweaking the model directly to improve its performance.
• The use of hash-anchored edits and other optimizations in Dirac can lead to significant reductions in API costs and improvements in code quality.
• There is a trade-off between the use of more expensive and advanced models, which can lead to better performance but also increase costs and reduce speed.