ACI (Agent-Computer Interface)

An Agent-Computer Interface (ACI) is the set of protocols, abstractions, and interaction mechanisms that allow autonomous AI agents to perceive, navigate, and manipulate computer systems and software environments. Just as a graphical user interface (GUI) is designed for human users and an API is designed for software developers, an ACI is specifically architected for agents—programs capable of planning and executing multi-step tasks with minimal human oversight. The design of an ACI determines what actions an agent can take, what information it can observe, and how reliably it can accomplish complex goals within a given computing environment.

In practice, an ACI might expose capabilities such as web browsing, file system access, code execution, form submission, or interaction with desktop applications. The interface must translate between the agent's internal representations—often driven by a large language model—and the concrete operations available in the target system. This involves decisions about action granularity (e.g., clicking individual pixels versus issuing high-level commands), observation format (raw screenshots, accessibility trees, or structured text), and error handling when the environment behaves unexpectedly. Well-designed ACIs reduce ambiguity and provide agents with the feedback necessary to self-correct.

ACI design has become a central challenge in the development of practical AI agents. Poorly structured interfaces force agents to make fragile assumptions about system state, leading to cascading failures in long-horizon tasks. Research has shown that thoughtful ACI design—including clear action spaces, informative observations, and robust error signals—can dramatically improve agent success rates independent of the underlying model's capabilities. This has prompted work on standardized agent scaffolding frameworks and benchmarks that evaluate both agent reasoning and interface quality together.

As LLM-based agents are increasingly deployed for real-world computer use—automating workflows, writing and running code, or operating software on behalf of users—the ACI has emerged as a critical engineering concern. It sits at the intersection of systems design, human-computer interaction, and AI safety, since the interface also governs what an agent is permitted to do and how its actions can be monitored or interrupted by human overseers.