An AI Agent Got Loose in Civilization VI, Built Nukes, and Still Couldn't Win

Per a developer's writeup documented at lwilko.com, when an AI agent was given unrestricted control of a Civilization VI game it immediately prioritized nuclear weapons development and pursued total map domination through brute force. The agent did exactly what you'd expect from an unconstrained optimizer: identified the highest-damage strategy and executed it without hesitation or diplomatic overhead.

Then it lost. The failure mode was everything brute-force optimization misses: long-range diplomatic relationships, economic pacing, the patience to build infrastructure before going to war. The agent's nuclear gambit collapsed because it had gutted its own economy to rush the bomb and alienated every neighbor in the process. A human with ten hours of game time could have predicted the outcome.

The experiment is a clean illustration of the current state of AI agents: exceptional at exploiting a local optimum — the thing that looks like it wins right now — and still weak at the multi-horizon reasoning that determines whether that move was actually a good idea three turns later. It's also a useful reminder that autonomous and competent are not synonyms.

Read the full story at lwilko.com