What is an agent-computer interface (ACI)?
An agent-computer interface (ACI) is the set of tool definitions, parameter formats, documentation, outputs and error messages through which an AI agent works with software, designed with the same care as a user interface for people.
Anthropic's "Building effective agents" (December 2024) argued that teams should put as much effort into agent-computer interfaces as they do into human-computer interfaces. The agent never sees the system behind a tool. It sees a name, a description, a parameter schema and whatever text comes back, and its decisions are only as good as those.
The post's evidence came from its own SWE-bench agent: the team spent more time improving tools than the overall prompt. When the model made mistakes with relative file paths after moving out of the root directory, the tool was changed to require absolute paths, and the model then used it without error.
What makes a good ACI
Formats the model already knows
Keep inputs and outputs close to text the model has seen widely, and avoid formats that need line counting or heavy string escaping.
Mistake-proof arguments
Change parameters so errors are hard to make, the poka-yoke principle Anthropic cites, such as requiring absolute paths.
Descriptions written like docstrings
Explain each tool and parameter the way a good docstring would for a new developer, including edge cases and when not to use it.
Fewer, consolidated tools
Anthropic's post on writing tools for agents (September 2025) favours a few tools built around workflows over one wrapper per API endpoint, with namespaced names.
Output worth reading
Return meaningful, human-readable fields rather than raw IDs, offer concise and detailed modes, and paginate or truncate long results.
Errors that steer
An error message should tell the agent what to try next, not just that something failed.
Examples
Anthropic reported that adding example calls to tool definitions raised accuracy on complex parameter handling from 72% to 90% in internal testing.
Testing an ACI
An ACI is tested the way a UI is: by watching it being used. Anthropic recommends running agents against realistic tasks, reading the transcripts to see where they stumble, and even letting an agent propose better tool descriptions, then measuring again.
How Motiontic's engineers apply it
On client systems, Motiontic's engineers rarely expose internal APIs to an agent as they are. They wrap them in a smaller set of task-shaped tools with plain names, readable output and guard rails in the arguments, then iterate on descriptions until eval transcripts show the agent choosing and using them correctly.
An API designed for developers compared with an agent-computer interface
| API for developers | Agent-computer interface | |
|---|---|---|
| Reader | A developer who reads the docs once | A model that reads the definitions on every call |
| Cost of verbosity | None at run time | Tokens on every request |
| Shape | One endpoint per resource or action | A few tools built around tasks |
| Output | Complete data, often IDs and codes | The fields the agent needs, in readable form |
| Errors | Status codes | Messages that say what to do next |
Common questions
Where does the term agent-computer interface come from?
Anthropic's "Building effective agents" (December 2024) popularised it, by analogy with human-computer interface, as one of three principles for building agents alongside simplicity and transparency.
Is MCP an agent-computer interface?
MCP is a protocol for exposing tools. The ACI is the quality of what gets exposed: the names, descriptions, parameters and outputs. A tool served over MCP can still have a poor ACI.
How is an agent-computer interface improved?
By running agents on real tasks, reading where they misuse a tool, and changing the tool's name, description, parameters or output, then measuring again with evals.
Related
Tool calling
How a model asks software to do something: a named tool, typed arguments, a result.
Model Context Protocol (MCP)
The open standard for plugging tools and data into AI agents.
Harness engineering
Designing the environment an agent works in, so its output can be checked and shipped.
Computer use agent
Agents that work through the screen for systems with no API.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.