Magic Tower AI Lab: the thesis in three minutes. Voice-over script 461 words over 180 s (about 154 words per minute). Each line starts at the time shown; the captions use the same text. [00:00] The question 00:00.4 Can a language model plan a route of seventeen decisions, when a single wrong one loses the floor? 00:05.8 This thesis tests that in Magic Tower, a puzzle game with no randomness and no undo. [00:12] Magic Tower 00:12.2 Every fight costs HP by one formula: attack lowers the number of strikes, defence lowers every strike. 00:19.0 So order matters: a gem taken before a fight saves more than its face value. 00:25.0 And nothing comes back: keys, HP and monsters are gone once used. [00:30] One decision 00:30.2 Here is the first decision on the simple floor. Four actions are legal. 00:37.0 The exact solver shows that only one, the Black Slime, still leaves a win. The other three have already lost the floor, though nothing shows it yet. 00:44.0 In the re-run, nineteen of twenty episodes chose it. [00:50] The hybrid agent and the exact solver 00:50.2 The agent splits the work: the engine computes reachability, exact damage and a lookahead; the model only picks a numbered action. 00:58.0 An exact solver scores every state. It also corrected a published reference: the complex floor's optimum is 68 HP, not 22. 01:04.2 Question one: what does the engine's observation contribute? [01:10] RQ1 · The observation 01:10.2 Failed episodes kept citing lines that showed only one side of a trade-off: a saving, without the HP it costs to get there. 01:18.0 Stating the missing side, with the model and the prompt unchanged, took the largest model from zero of five wins to ten of ten. 01:26.2 Across both runs it won fifteen of fifteen, though always at 2 HP, where the optimum is 13. [01:35] Two warnings and the boundary 01:35.2 Two warnings. Every early win had contained a silent engine override; overrides are now mechanical only, and logged. 01:43.0 And on the complex floor no model wins: twenty-nine of thirty losses fall at one cheap fight, long before the door a human would blame. [01:50] RQ2 · Prompt optimisation 01:50.2 Question two: does evolutionary prompt optimisation help? With the annotations present, its gains stay within the noise between two runs of one prompt. 01:59.0 With a reasoning model, both search loops raised their training score and lost on held-out floors: the population search won 24 of 39 games, the plain rules prompt 32. [02:08] RQ3 · The planning horizon 02:08.2 Question three: react at every step, or commit to a plan? 02:16.0 On the simple floor, reactive play won fourteen of twenty episodes; committed plans one of seventy. Plans match it with a fraction of the calls only when the order is given. [02:25] The 24-floor tower 02:25.2 On the full 24-floor tower, the rule-based heuristic stops at floor 8. The best model reached floor 18. 02:33.0 Resumed there with blue keys for sale, as in the original game, it reached the normal ending. [02:40] Takeaways and this site 02:40.2 The lesson: put strategic knowledge into the engine's observation, with both sides of every trade-off, not into the prompt. 02:49.0 Commit to a plan only when the order is known, and measure agents against an exact reference. 02:55.0 Every episode, table and search is on this site. Try them.