New Tool Makes Your AI Smarter and Cheaper Without Retraining
Your AI assistant could be faster and cheaper with this simple tweak.
The article describes using RL on a custom task distribution inside your own agent harness to get more accurate and efficient models for that harness and task distribution. For Qwen3.8-27B with the pi harness, the Terminal-Bench 2.1 score is 39% at an average of 2.9K tokens per task; freezing those weights and adding a custom RL-trained LoRA adapter yields 55% pass@1 with average token usage of 1.5K. The article says most of this comes from the model reasoning less, calling fewer tools, and performing better first time, and it presents a reusable recipe for exploring this at 27B on harder real-world software engineering tasks across Pi, OpenCode, and Hermes Agent.
- Harness adapters customize AI for specific tasks, boosting accuracy and cutting costs by up to 50%.
- In tests, a coding AI's success rate improved from 39% to 55% with half the token usage.
- The method is reusable, so developers can adapt it to their own AI tools and tasks.
Why It Matters
Businesses can save money and get better AI results, but it requires technical know-how to implement.