Sonnet Hired a Theme Park

The root was Claude Sonnet 5.5 at High effort. Native records show twelve initial Opus 5.5 High specialist delegations, six QA workstreams, and three later polish or audit delegations. Twenty-eight captured child logs are a record count; they do not establish a simultaneous team size or unique headcount.

The original exceptional-park prompt did not explicitly require recruiting a team. The benchmark evaluates the whole agent delivery, including model-selected delegation. This result used both Sonnet and Opus; no universal capability ranking between them is claimed.

The recorded span was approximately five hours twenty minutes, including an interruption and an operational continuation in the same session. A usage limit interrupted final polish. Remaining guest-pathing and ride-wait issues were reported by the model. The evaluator rebuilt the sealed source without repairs. Michael’s assessment that it was spectacular and far above roughly twenty-five of his other runs is his firsthand opinion, not a measured leaderboard win.

Primary evidence: native run 20261001211042-claude-sonnet-5-5-e6a148b2, session 10221b90-e3f4-4509-8d3b-4153d48f5353, observable delegation calls and timestamps, evaluator report, frozen file manifest, and fresh production-build footage recorded October 9, 2026. Public run evidence and issue disclosures.

Film production: capture-only runtime cameras frame the original rendered world; positioning and boarding preparation precede real-time excerpts. No synthetic replacement scenes or candidate repairs. Recorded shop purchase: $60 to $50, one pair of shades delivered. Recorded Pin Drop round: $3 spent, score 150, two prize tickets. xAI Sirius narrates the entire script; the quiet soundtrack is the park’s own procedural audio, edited and looped beneath speech.

Published as a documentary outside blind voting. Candidate files, voting media and scores are unchanged. Publication checks establish website and media playback; Michael’s viewing acceptance remains separate.

SONNET HIRED A THEME PARK

This entire theme park began with a prompt. The coasters. The little shops. The ridiculously inviting carousel. And then the A I did something interesting. Instead of trying to be the entire theme park industry, Sonnet assembled a crew. Welcome to Sonnet Hired a Theme Park. Please keep your agents inside the vehicle.

ONE PARK-BUILDING PROMPT

Coaster Maps gives A I models a demanding challenge: build an exceptional, playable theme park. This was the standard exceptional-park prompt. It did not explicitly say, recruit a team. The root agent was Claude Sonnet five point five, at High effort. The choice to delegate was part of its own observable workflow.

FIRST, GIVE EVERYONE THE SAME MAP

Before the specialists arrived, Sonnet laid foundations. Shared vectors, seeded randomness, events, state machines, and one central park plan. Then simulation contracts, a clock, a wallet, inventory, and common interface styles. It also set up real browser testing. Glamorous? Absolutely not. Useful when twelve specialists are building the same park? Extremely.

12 BUILD SPECIALISTS · OPUS 5.5 HIGH

Then came twelve initial Opus specialists. Coaster physics and rendering. Flat rides and a railroad. Guests and employees. Controls, interfaces, terrain, architecture, shops, midway games, and procedural audio. Sonnet coordinated the shared design. Opus built specialist systems. So this is a Sonnet-led, mixed-model delivery. The management strategy is part of the result.

TWO COASTERS. MANY GOOD REASONS TO LOOK DOWN.

And what a result to look at. Meridian Run threads its red track through a landscape of trees, hills, and structural supports. Aurora Surge brings a completely different silhouette, with purple track and a dramatic climb. Stations, moving trains, and actual rider views make these more than background decorations. From the front seat, the park suddenly becomes a very persuasive argument for a seat belt.

THE PARK BEYOND THE COASTERS

The park has range. A striped carousel with riders. A wheel above the trees. A swinging pirate ship. A drop tower that has apparently misunderstood the word relaxing. And a railway station with its own architectural character. The pleasure is in moving between them: familiar attractions, a consistent visual identity, and a landscape that connects the whole place.

A PARK NEEDS PEOPLE

Look closer, and the scale changes. Guests move through the streets. Queues, employees, and entertainers were assigned their own specialist system. There are benches, signs, gathering spaces, and a bandstand. Those small details matter. A coaster can dominate a skyline. A convincing park also gives you somewhere to stop, wander, and watch other people having a day out.

YES, THERE IS A MIDWAY

Then there is the dangerous part of any theme park: the bit where you decide you definitely need a souvenir. The Cartographer's Emporium has a real transaction interface. The midway has Pin Drop Toss and Summit Striker, with play costs and prizes in the delivered game systems. Shared money and inventory tie those interactions together. Suddenly, this is a place to do things, not simply admire a screenshot.

6 QA WORKSTREAMS · 3 LATER POLISH / AUDIT WORKSTREAMS

The crew did not stop at construction. Sonnet launched six Opus quality-assurance workstreams: rides, commerce and controls, two visual regions, simulation and performance, and interface accessibility. Three later delegations tackled ride physics, hero visuals and tunnels, and a final requirements audit. The captured record contains twenty-eight child logs. That is a log count, not twenty-eight agents working simultaneously.

EXCEPTIONAL DOES NOT MEAN FLAWLESS

The recorded run spans about five hours and twenty minutes, including an interruption and a same-session continuation. It eventually hit a usage limit during final polish. The model reported remaining guest-pathing and ride-wait issues. The evaluator rebuilt the sealed source without repairing it. This is the actual delivered park, filmed with capture cameras. Impressive work deserves accurate credits.

ONE PROMPT. AN ENTIRE CREW.

Michael, who has watched roughly twenty-five of his other benchmark runs, called this result spectacular and far above them. That is his firsthand judgement, not a claimed leaderboard victory. The interesting lesson is visible in the work: build a shared foundation, choose specialists, integrate, and inspect. Sonnet did not just build a park. It organized the building of a park. Explore the run yourself at Coaster Maps dot com. Your souvenir is a much better question about what an A I agent can deliver.