AI Agents Learn to Code Faster with New Training System

Training AI agents to handle complex, multi-step tasks like software engineering has long been a costly and inefficient process, often requiring massive computational resources and ad-hoc solutions. Researchers have now developed a new framework that dramatically improves this efficiency, enabling agents to learn faster and perform better with significantly reduced costs. This advancement could accelerate the deployment of AI assistants in real-world scenarios, from coding and web browsing to deep research, by making agent training more scalable and accessible.

The key finding from this research is the creation of S KY RL-AGENT, a framework that trains AI agents more efficiently through optimized scheduling and tool integration. Using this system, the team trained SA-SWE-32B, a software engineering agent based on the Qwen3-32B model, which achieved a 39.4% Pass@1 rate on the SWE-Bench Verified benchmark. This performance matches state-of-the-art models of similar scale but with more than a 2× reduction in training cost, demonstrating that smarter training s can yield high-quality without exorbitant expenses. The agent also showed improved generalization to other tasks like terminal command execution and web navigation, indicating that the skills learned are broadly applicable.

Ology behind S KY RL-AGENT involves three main components: a tool-centric agent loop for flexible integration of tools like code editors and search utilities, a fine-grained dispatcher for efficient scheduling of tasks across CPU and GPU resources, and a backend bridge that connects to existing reinforcement learning frameworks. Specifically, the framework uses an asynchronous pipeline dispatcher that overlaps CPU-bound operations (like runtime initialization) with GPU-bound inference, achieving a 1.55× speedup over naive s. This is complemented by a tool-enhanced training recipe that includes an AST-based search tool to help agents navigate codebases more effectively, boosting sample efficiency and rollout success rates.

From the paper show clear improvements in both performance and efficiency. As illustrated in Figure 1a, training metrics for SA-SWE-32B, such as reward and average number of turns, improved steadily over 125 steps, with the non-resolved rate remaining lower than baselines like DeepSWE. Figure 1b demonstrates that the async pipeline maintained GPU utilization at around 90%, avoiding the fluctuations seen with async batch strategies. On benchmarks, SA-SWE-32B outperformed the base Qwen3-32B model, with scores increasing from 13.75 to 16.25 on Terminal-Bench and from 3.68 to 4.6 on WebArena, as shown in Table 3. These gains highlight how the framework's optimizations translate into tangible benefits across diverse agentic tasks.

Of this work are significant for both AI research and practical applications. By reducing training costs and improving efficiency, S KY RL-AGENT makes it more feasible to develop AI agents for real-world use, such as automating software bug fixes, assisting with web-based tasks, or conducting deep research. The framework's modular design allows it to be easily adapted to different domains, as demonstrated with case studies on deep research, computer use, and memory agents. This flexibility could lead to broader adoption in industries where multi-step, interactive AI systems are needed, potentially speeding up innovation and reducing reliance on expensive, proprietary solutions.

Despite these advances, the paper acknowledges several limitations. For instance, the computer use agent trained with S KY RL-AGENT showed improved training rewards but little gain in validation accuracy, suggesting that tasks like GUI interaction remain challenging for current models to generalize. Additionally, the deep research agent faced issues with tool overhead and potential benchmark answer leakage from online searches, requiring domain blocks to ensure fair evaluation. These s indicate that while the framework improves efficiency, further work is needed to enhance robustness and generalization across more complex and varied environments.