Training AI agents to handle complex, long-term tasks is notoriously expensive and difficult. Human experts can’t scale, and standard AI tools often break the delicate chain of logic required for these workflows. Now, researchers have introduced a method that automates this process, creating thousands of high-quality training tasks at a fraction of the cost.

The new framework, called Recursive Synthetic Terminal Tasks, works like an infinite loop of creation and verification. It starts with a few verified seed tasks and recursively expands them. The system extends the solution, realigns the instructions and verification tools, and tests the result in a fresh environment. If it passes, the new task becomes a seed for the next round. This cycle repeats, allowing the system to generate complex tasks without human intervention.

The results are striking. After fifteen rounds, the framework produced 37,484 tasks at roughly $0.05 each. The difficulty skyrocketed, with the median solution growing from 67 to 374 lines of code. Consequently, the success rate of DeepSeek-V4-Pro on pass@4 plummeted from 90% to 2.5%. This proves the tasks are genuinely challenging and not just easy variations of the originals.

Training models on these synthetic tasks yielded massive improvements. Fine-tuning improved performance by up to 10 points on Terminal-Bench, Terminal-Bench Hard, and Long-Horizon Terminal Bench. When combined with agentic PPO, gains reached up to 41.2% on hard tasks. Crucially, the process showed no signs of slowing down, suggesting this method can keep generating harder, better data indefinitely.