Tour · 11 of 11 · Sections 6.6–6.7, 7.5 (RQ3)

Planning and the tower

Do longer plans help, and how far up the 24-floor tower does the agent get?

The title asks about a move from reactive control to holistic planning. The agent can decide one action per call, or commit to k actions, to a plan for the whole floor, or to a single plan made once. Committed plans need fewer model calls.

14 / 20reactive wins on the simple floor, Qwen3.5-397B and DeepSeek-V4-Flash

1 / 70committed-plan wins, same models and floor, including plans of self-chosen length [0.00, 0.08]Tables 6.9–6.10

¼ – ⅑of the calls for the same success, only where the solver's route is in the prompt

A plan step is chosen before the engine has computed the facts of the state it leads to. Where the order must be derived from those facts, committed plans fail; where the order is given, they transcribe it a few steps ahead at lower cost.

The tower is the long version of the problem: 24 floors, shops, keys carried between floors, and the normal ending on MT21. The highest floor reached rose across three tower observations; the heuristic stops on MT8.

heuristicfirst observationsecondfinal

Rule-based heuristicMT8

Qwen3.5-397BfirstMT7

Qwen3.5-397BsecondMT8

Qwen3.5-397Bfinal, best of 3MT18

DeepSeek-V4-Flashfinal, best of 3MT17

Qwen3.6-35Bfinal, best of 3MT16

Mistral Medium 3.5final, best of 3MT8

Qwen3.5-397B, resumed on MT18blue keys for saleMT21 normal ending

Highest floor reached (Table 6.13); the green mark is MT21, the normal ending. All tower runs.

What it adds up to. The agent's decisions were rarely better than what the engine stated at the moment of decision. When the engine stated both sides of every trade-off, Qwen3.5-397B won the simple floor in 15 of 15 episodes and the prompt mattered little. When it did not, neither a better prompt nor a longer plan recovered the loss.

Continue with the results, the episodes or the comparison of a win and a loss, or play the simple floor yourself.