How do long-horizon RL environments actually scale coding agents? In Part 2, pre.dev Labs breaks down custom GSPO, treating SFT as a rescue operation, and how catching bad signal early cut training costs by 5x–7x while quadrupling Terminal-Bench 3.0 scores.

Do long-horizon RL tasks move the needle for coding agents? pre.dev Labs trained Qwen3.8-27B on ~100 frontier tasks, boosting its Terminal-Bench 3.0 score from 1/74 to 4/74. Here is how we built the environments and scaled agent performance.

Stop burning tokens on brute force. If you’re using coding agents but seeing diminished ROI, this breakdown is for you.