When routine operations break architectural pattern relationships in LLM Agent Harnesses

Researchers propose the PatternOps evaluation method and find that routine operations such as reapplying a plugin or saving an allow rule can silently break architectural pattern relationships such as guardrails in LLM Agent Harnesses without any code changes. Such operations cause the policy to stop evaluating edits while all plugins remain active, allowing a prohibited edit to be executed. The study covers five relationships related to guardrail enforcement points, autonomy controllers, and the blackboard pattern across five published harnesses, with a main sample of fourteen independently written plugins. Assignments that preserve the required order or retain and rebuild the required state hold on their test paths and operation sequences; other assignments, although they prevent the prohibited effect, miss edits that are rejected elsewhere. The paper notes that in these harnesses, pattern relationships are properties of the running system and its history, and that architecture designers should evaluate them in light of the decision paths, state lifecycles, and operation sequences actually used in deployment.