ML//agent//agent-computer interface

An **Agent-Computer Interface (ACI)** is the set of commands and feedback formats through which a language-model agent operates a computer. It plays the role that a GUI or shell plays for a human, but it is designed around a model's ability to generate structured actions and interpret serialized observations. The term was introduced by [SWE-agent](https://arxiv.org/abs/2405.15793) in 2024.

An Agent-Computer Interface (ACI) is the set of commands and feedback formats through which a language-model agent operates a computer. It plays the role that a GUI or shell plays for a human, but it is designed around a model's ability to generate structured actions and interpret serialized observations. The term was introduced by SWE-agent in 2024.

On the action side, names, schemas, editing primitives and validation rules determine which operations the model can express reliably.

On the observation side, file viewers, search summaries, command feedback and errors determine which distinctions the model can perceive, that is, how computer state becomes model-readable evidence (observation design).

The SWE-agent experiments showed how much these choices weigh: a 100-line file view beat larger windows, listing only the files containing search matches beat richer match output, and an explicit success message removed ambiguity when a command returned nothing.

The ACI is the exposed interaction surface. The agent harness is the wider runtime that implements and governs it, and owns context assembly, execution, permissions and lifecycle.

Giving two agents access to the same computer does not give them the same usable computer, because interface design changes the task they can effectively solve.