Should an AI describe its world in its own words, or fill in a fixed form, when every world is different?

This explores why letting an AI describe the state of its world in free-form terms (an open schema) might beat forcing it into predefined fields (a fixed schema) when the environments vary a lot. The corpus has no note that runs this exact head-to-head, so the evidence below is indirect. It points one way, and the case for open schemas is a well-motivated hypothesis here, not a measured result.

The closest evidence is a world model that keeps its state in plain language. Qwen-AgentWorld predicts the next state of an environment in natural language, trained on more than 10 million agent trajectories. It beat training in the real environments on three benchmarks and transferred across seven domains (Can language models learn to simulate agent environments?). The note doesn't say the format caused the transfer. Still, one representation covering seven different worlds is what a fixed schema struggles to do, because a web page, a shell session and a game each need different fields. An open format means nobody has to decide up front which fields matter.

The corpus also shows why deciding up front is risky. A state whose shape is fixed before the data arrives can't hold what it didn't anticipate. Transformers provably beat state-space models at copying and retrieving from context because state-space models squeeze everything into a fixed-size state, while transformers keep the raw context to query later (Can state-space models match transformers at copying and retrieval?). The same limit shows up in safety. Per-action checks can't state a rule like "these two individually fine actions are not fine together," because their state has no place to record history (Can stateless checks ever catch sequence-level constraint violations?). Whatever the state format leaves out, you can't ask about later.

The complication is that structure isn't the villain. DeepAgent folds its history into episodic, working and tool memory schemas, and that structure saves tokens and lets the agent pause and rethink. What prevents degradation is that the agent chooses when and what to fold, so structure and autonomy work together (Can agents compress their own memory without losing critical details?). The problem is a designer freezing in advance what counts as important. Two other notes show the alternative of deciding what matters when the question arrives. One is an agent that reconstructs memory by walking a graph guided by evidence, gaining up to 23% over fixed retrieval pipelines (Can agents reconstruct memory on demand instead of retrieving it?). The other is a layered harness that keeps disk-backed history addressable so later steps can pull out what they need (Can external state caches let models solve harder problems?).

The pattern is that an open schema postpones the question of what matters until the world shows up, which is why it fits diverse worlds. The corpus doesn't measure the cost: free-form state is harder to check, compare or constrain than a typed field. Whether the flexibility is worth that is still untested in these notes.