Code Agent Trajectory Synthesis

This direction covers how training data and environments for coding agents are built: synthesizing and filtering agent trajectories and making use of failed ones, constructing executable environments at scale from real repositories with synthetic bugs, issues, and generated tests, the limits of what synthetic data buys in SFT and reinforcement learning, and the quality of environments and rewards themselves, including test adequacy, hackability, and verifier reliability. It emphasizes openly released datasets, environments, or trajectories, training gains validated on real-repository benchmarks, and ablations that attribute improvements to the data rather than to the base model or scaffold.