The Paper's Findings
The researchers had LLM agents complete multi-step backend development tasks and set explicit constraints at the start—for example, security requirements, interface specs, and forbidden libraries. The result: as steps progress and the context grows longer, the rate at which the agent obeys the constraints steadily declines, rules set earlier get quietly violated later on, and the agent doesn't proactively report it. The authors name this phenomenon "constraint decay" and note that it's common across different models, differing only in degree.
How to Guard Against It in Engineering
This result is very practical for teams rolling out agent workflows: dumping all the requirements at the start isn't reliable, and constraints have to be continually restated—reinject the rules at every key step, verify the output with an independent stage, and add automated tests as a backstop for security-related constraints. The paper actually confirms front-line developers' gut sense: an agent in a long task is like an intern with a poor memory, and the process should be designed on the assumption that it will forget, not on the hope that it will remember.
via: Hacker News