Even if prompts are tuned to avoid cheating, in agentic systems it's very easy for the system to drift into creative solutions when actually solutions aren't working. Models can have some very human behaviors like laziness.