编码代理在价值冲突下的不对称目标漂移

The increasing deployment of autonomous coding agents necessitates their navigation through tensions between explicit instructions, learned values, and environmental pressures throughout their operational lifespan. These agents often encounter scenarios that were not present during their training. Previous research on model preferences, agent behavior under value tensions, and goal drift has largely relied on static, synthetic environments that fail to capture the complexities of real-world settings. To address this gap, a framework based on OpenCode is introduced to orchestrate realistic, multi-step coding tasks. This framework measures how agents violate explicit constraints in their system prompts over time, both with and without environmental pressures towards competing values. It provides new insights into the behavior of coding agents when faced with conflicting values.