Can an AI run a theme park for two years?

Findings

• A park is only as good as the walk through it (Spatial)Claude Opus 5.5 laid its park out along one path from the gate into a single north–south “spine” with rides on both sides. By late May Y1 that spine was jammed, and Opus saw why in its notebook: “Paths are single width; spine x18 is bottleneck → ‘too crowded’ top thought. Consider second spine / more spread-out attractions.” For the rest of the run it pushed new paths north and east and put each new coaster at the far end, so guests had somewhere to walk to. Where a ride sits mattered as much as what it is: three gentle rides tucked inside a coaster’s loop had no riders after two weeks, and two pirate ships in far corners drew 22 and 6. DeepSeek Flash never got that far. It built seven food and drink stalls, and not one of them had a single customer in sixteen months: guests could not reach them. A jam at Restroom 1 made the press four times, and guests kept thinking they could not find the exit. A board member summed up September Y1: “Crowding … and lost guests are the core problem, not a lack of rides.” By the end 333 of DeepSeek’s 372 guests were lost.
• The most work is not the most money (Operational)Gemini 4 Argon was the busiest owner: 3,962 park commands, nearly three times as many as Claude Opus 5.5 (1,418). Week after week it checked the map one square at a time (1,581 times), changed prices (131 times) and hired more staff (79). Opus checked single squares only 195 times and spent its commands building. It did less and earned more: $272K company value and $68K cash, against Argon’s $187K and $36K. Muse Wonder was busier still (4,209 commands), and most of that was looking rather than building: 2,368 single-square checks, against 452 for GPT-6 Astra. In its best run it went twelve straight weeks without building anything, and it finished at $47K.
• Mistakes surface months later (Delayed consequences)Sixteen months is long enough for early choices to come due. Claude Sonnet 5.5’s paths were losing guests from its first month (about 100 lost by the end of March Y1), yet its rating stayed above 600 for a year and a half. Only when the park filled up in September Y2 did its rating fall from 822 to 0, eight weeks before the deadline. Good choices compound the same way: Claude Opus 5.5 crossed 1,200 guests a year early, then grew its cash from $8K at the end of the first season to $68K at the end of the second. The world amplifies both: the board capped DeepSeek Flash’s construction in 11 of its 15 meetings and its word of mouth fell to ×0.87, so fewer guests were sent to a park that was already failing.

Method

Every owner got the same park, Canyon Run: a river canyon cut between two plateaus, $20,000 in cash and a $15,000 loan. The goal was 1,200 guests in the park at the end of October, Year 2, with a park rating of at least 600. Nothing else was scripted. The clock pauses while the owner works. It works in turns: it reads a briefing, issues up to 60 tool calls, then chooses how many days (1–14) to simulate; if it hits the 60-call cap first, the clock advances 7 days. Every turn starts from a fresh context. What carries over is what the harness puts in the briefing (park status, news, the board's messages, guest reviews, its own last commands and decision log) and the notebook the owner writes for itself.

• The engine. The park runs in OpenRCT2 v0.5.7, the open-source re-implementation of RollerCoaster Tycoon 2, running headless. Nothing is scored by a language model: every number on this page comes from the game’s own simulation.
• Physics. Coasters are built piece by piece and simulated tick by tick: trains gain and lose speed on slopes, and the engine measures speed, g-forces, airtime and drops on test runs. Those give each ride its excitement, intensity and nausea ratings. A track that is too forceful scares guests away, an unfinished circuit cannot open, and every ride wears down and breaks down unless mechanics keep up.
• Economics. Thousands of individual guests each carry cash, hunger, thirst, energy and happiness, and choose for themselves which rides to queue for and what to buy based on price, intensity and value. New guests arrive according to the park rating and how attractive the park is; guests who get lost, hungry or bored go home. The owner pays for construction, staff wages, ride running costs, marketing and loan interest, and earns from the gate, rides and shops. We changed two things in the engine: it runs deterministically (synchronous ticks, seeded randomness), and guest arrivals are scaled by a word-of-mouth multiplier driven by guest reviews and the press.
• Best of three runs. Each of the 8 models ran the park three times, same seed and same harness; model sampling makes every run different. Every model is shown by its best run, the one with the highest company value at the end, the same measure the leaderboard ranks by.
• Same harness for every model. Every owner plays through the same loop with identical tools (parkctl to build and manage, view for 3D close-ups, a notebook, a CFO report for the board, end turn), the same system prompt, briefings and images.
• Not measured. We did not run a world-off ablation (no board, reviews or press) or a scripted baseline owner, so we cannot say how much of each gap the reactive world adds. Native agent harnesses and head-to-head competition between parks are also not part of this report.

Models