ML//agent//agent harness//model-relative interface design

An agent interface is **model-relative** when its usefulness depends on the model and post-training policy consuming it. A schema can be valid for every MCP client while remaining cognitively ergonomic for one model, merely usable for another and actively confusing for a third.

An agent interface is model-relative when its usefulness depends on the model and post-training policy consuming it. A schema can be valid for every MCP client while remaining cognitively ergonomic for one model, merely usable for another and actively confusing for a third.

NamesAnthropic found that prefix and suffix namespacing have non-trivial effects that vary by LLM.

FormatsJSON, XML, Markdown and prose can produce different outcomes because models inherit different training distributions and parsing habits.

GranularityStronger models may exploit general shell tools; other models benefit from narrower operations that encode more workflow.

BudgetA coarse tool can dominate with five calls remaining while fine-grained tools preserve flexibility under a larger budget.

VersioningChanging the model can invalidate interface optimizations even when every software type and endpoint remains compatible, because the hidden ABI holds only while the target model still reads the interface as intended.

PortabilitySyntax portability does not imply cognitive portability, so interoperability claims need cross-model behavioral evaluations.

The practical consequence is co-adaptation: model and harness should be evaluated as a pair, and old scaffolding should be stress-tested whenever model capability changes. The same holds for comparing harnesses: two of them can be equivalent for one model and observably different for another (harness equivalence).