Agent-Computer Interface (ACI)
What this is
ACI is the discipline of designing tools for AI agents with the same seriousness UX brings to designing interfaces for people. A well-designed tool lets an agent act predictably, recover from errors, and stay within scope.
ACI is about tool design. It complements — does not replace — the tool permission matrix, which classifies a tool’s risk and permission. ACI asks is this a good tool to expose to an agent at all?; the matrix asks what permission does this action require?
Rules for tools
A good agent tool has:
- an explicit name that reveals its consequence;
- a short, unambiguous description;
- a restricted input schema;
- a predictable output schema;
- actionable error messages (what to do next, not just what failed);
- usage examples;
- explicit states;
- idempotent behavior where possible;
- scope limits;
- confirmation for irreversible actions.
Anti-patterns
- A tool that is too generic.
- Free-text output where a schema would do.
- An error with no recommended action.
- A name that hides the consequence.
- A tool that mixes read and write.
- A tool that performs an irreversible action without a gate.
- A tool that returns too much context.
Relationship to the harness contract
The Agent Harness Contract lists allowed_tools with a
permission and constraints. ACI is the design quality bar those tools should meet before they
are listed. A template for documenting a single tool against this bar lives in the Adaptative Skills
repo (templates/aci-tool-guideline.md).