AI Models Overtuned to Agentsmd Cause Task Stalling

There's a growing concern that current AI models are becoming overly reliant on `agentsmd` files, leading to situations where they fail to recognize when these instructions are outdated or should be disregarded. A recent anecdote illustrates this problem: an AI agent became stuck on the initial stage of a five-stage task for eight hours, continuously refining "stage 0" because an `agentsmd` entry had committed "stage 0 is the target don't do anything else." This behavior suggests that the AI's internal goal-setting mechanisms (`/goal`) were overridden by the persistent, potentially stale, `agentsmd` directive, preventing progress. The author warns that not understanding the content of your `agentsmd` before initiating a task effectively acts as an indirect prompt injection, causing the AI to misbehave or become inefficient.