MrBeast Super Bowl Puzzle — a lived multi-agent commission (founder case study)

Primary-source account from the founder (iMessage 2026-06-09), captured faithfully. Surfaced inside the "commissioning" thread (see [[2026-06-09-claude-md-prompt-precedence-full]] is unrelated; see the commissioning brief this feeds). This is the founder's own early experience of running a multi-agent swarm against a vague, outcome-defined objective — possibly predating the creation of "Ray" (he wasn't sure if Ray existed yet).

The commission

Objective (the only fixed requirement): get the code (solve the MrBeast Super Bowl puzzle).
Everything else: open-ended. No prescribed method. Pure "churn until we find a solution."

The swarm (5 roles)

Community-scouring agent — continuously scanned the community for what others had discovered publicly.
Browser agent — submitted candidate answers to "beast bot" to gather feedback and clues from its responses.
Video agent — dissected the set of MrBeast videos frame by frame looking for embedded clues.
Puzzle solver — worked on cracking the clues the others surfaced.
Coordinator — kept the current state and who-was-responsible-for-what up to date. (The shared-state / blackboard hub.)

Outcome

Did not win. But the founder's reflection: "that sort of vague requirement and churn until we find a solution sure was fun and unique to working with agents."

Why this is in the vault

It's a concrete, founder-lived instance of the two patterns the commissioning brief is about:

Outcome-defined commission — specify the finish line ("get the code"), stay fully method-agnostic, let the makers churn. Same family as the Longitude Prize / XPrize "design the finish line, not the method" principle.
Coordinator + shared-state multi-agent architecture — independent specialist agents (search, browser/feedback, perception, reasoning) writing into a coordinator that holds global state and assigns responsibility. This is the classic blackboard/coordinator orchestration pattern, and it directly prefigures how Ray now runs the Workflow tool (pipeline/parallel fan-out with a synthesizing seat).

Mapping against Ray Data Co

Validates the founder's instinct that the highest-energy, most "fun and unique" agent work is the vague-requirement, churn-to-solution mode — which is exactly the "commission, don't steer" thesis from the Fable 5 / Mollick thread ([[2026-06-09-fable5-workflow-optimization-memo]]).
The coordinator role he built by hand is what the Workflow tool now provides structurally (deterministic JS orchestration, a synthesis seat holding aggregated state). Worth asking: where else would a persistent coordinator-with-shared-state beat a one-shot fan-out?
Candidate anchor case study for the forthcoming commissioning → agent-spec brief.

[[2026-06-09-fable5-workflow-optimization-memo]] — the "commission a studio, sign off on the work" thesis
[[project_l5_north_star_strategic_direction]] — agent-capability-first direction