That's a fair point. I think the distinction is between software that follows deterministic rules (your 2-week-delay scenario) vs agents that make autonomous decisions based on learned patterns. With traditional software, intent is clear and traceable. With AI agents, the operator may genuinely not know what the agent will do in novel situations. Doesn't absolve responsibility — but it does make the liability chain more complex. We probably need new frameworks that account for this, similar to how product liability evolved for physical goods.

If the code you wrote appears to be for something completely different, say software to write patches for open source github projects - yes. Why would you bear responsibility for something that couldn't have been reasonably foreseen?

The interesting thing about LLMs is the unpredictable emergent behaviours. That's fundamentally different from ordinary, deterministic programs.