New Training Method Makes AI Agents More Reliable at Database Tasks

agents (software that can take actions like booking a hotel or updating a support ticket) often say they're done when they've actually made mistakes. A new open-source tool called ThinkingBox fixes this by checking the final database records—not just what the AI says—to see if the job was truly completed. It then uses that same check as a reward to train the AI to do better.

The team trained a mid-sized AI model (Qwen3.8-27B) on 507 realistic tasks. Before training, it succeeded on average 52% of the time and perfectly completed only 38 out of 507 tasks on all 20 attempts. After just 50 training steps, average success rose to 61%, and the number of perfectly completed tasks jumped to 107. In one example, the AI learned to stop changing a hotel booking without required manager approval.

This matters because businesses want AI that can handle back-office work without constant human supervision. The training method focuses on tasks that are challenging but not impossible for the AI, making learning efficient. The tool is now available on GitHub for anyone to use.

The catch: the AI's success rate on solving each task at least once dropped slightly, meaning it became more consistent but slightly less able to solve the hardest tasks. Also, the training requires a database setup and careful task selection, so it's not plug-and-play for all situations.