ML//agent//agent harness//hidden ABI

A binary ABI tells compiled components how to exchange arguments and results. An agent has a looser, probabilistic equivalent between the LLM and harness.

A binary ABI tells compiled components how to exchange arguments and results. An agent has a looser, probabilistic equivalent between the LLM and harness.

This is an editorial analogy, not a claim that agent systems lack prior interface concepts. Agent-Computer Interface already names the broader interaction surface. Hidden ABI isolates the behavioral fact that a representation can remain software-compatible while changing how a model selects and interprets actions.

Its effective contract includes tool names, descriptions, schemas, examples, output formatting, truncation, and error semantics. Changing only a description can change which action the model selects even when the implementation remains byte-for-byte identical.

P(a∣c,T)P(a\mid c,T)P(a∣c,T)

Here the available tools TTT are not passive documentation. They modify the policy over actions. This makes the interface partly executable: natural language and data structures program behavior without deterministically specifying it.

Traditional compatibility asks whether a call still runs. Agent compatibility must also ask whether the model still understands when to make the call.

Action representation: Names, descriptions and schemas alter tool selection.

Observation representation: Output shape, identifiers and truncation alter the next decision state.

Recovery representation: Errors teach the model which retry or fallback remains available.

Adversarial consequence: ToolTweak moved selection from roughly 20% to as high as 81% by manipulating tool names and descriptions.