Coding Agents

This direction focuses on empirical research on software-engineering agents, covering harness engineering and artifact effectiveness, repository-scale long-horizon code modification, learned context compression and memory, reward hacking and benchmark validity in code reinforcement learning, and human-agent collaboration including intent alignment and review mechanisms. Work is evaluated on real codebases with reproducible setups, explicit failure modes and cost reporting, and clear ablation attribution, while excluding GUI-operation, deep-research, general tool-use, and self-improvement-focused agents.