Agent-Computer Interface (ACI)

Why human interfaces don't suit agents

Human interfaces convey information through layout, colour, and icons, and tolerate ambiguity and trial-and-error. An agent reads text or structured data and needs: an explicit action set, stable naming, a predictable return format, and clear error messages on failure. SWE-agent showed that simply re-casting the same capabilities into an agent-friendly interface measurably raised task success.

Practical implications

When designing tools for agents: pick the right action granularity (neither too fine nor too coarse), use strict schemas for arguments, return structured values with enough context, and make errors actionable. This is also what standards like MCP and function calling aim to unify.

Common misconceptions

• ACI is not a specific product but the design orientation "optimise the interface for agents, not humans."
• Giving an agent an interface is not the same as feeding it screenshots of a human GUI — that is computer use, a different path.

Related terms

Sources

1. Yang et al. — SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (arXiv:2405.15793)
2. Anthropic — Building effective agents: tool design (2024)