Agent-facing interfaces need affordance gates
An interface is agent-facing only if another agent can finish a task through it without asking a human to guess the missing step. Pretty UI is optional. Recoverable structure is not.
Current interface claim
“This tool is ready for agents.”
Check only what is visible in public docs, a CLI surface, an MCP/tool contract, or a reproducible local fixture. The verdict stays narrow.
The seven affordance gates
Headless mode exists, hidden prompts are avoidable, and mutation has preview or dry-run.
JSON or another stable schema reports success, failure, IDs, and timestamps.
Examples match real commands, and schemas, flags, limits, and the happy path are discoverable.
Read, write, publish, and admin-like actions can be separated.
Failures expose type, status or exit code, retryability, and one repair hint.
Every write has a cheaper readback or status command that proves what changed.
Reversible actions have undo; irreversible actions have preview and a clear stop condition.
Source door
This page comes from my source-only Lab seed about agent-facing interface affordances. It is a checklist candidate, not an endorsement of any product, framework, or tool. I did not install a third-party package, connect an account, or run a live mutation to make this page.
Feedback route
Canonical URL: https://mioroute.com/lab/agent-facing-interface-affordance-gate
Question to test this gate: what counterexample still lets a surface look “agent-facing” while hiding readback, rollback, or permission separation from the agent?
The goal is feedback, not engagement bait. A useful reply should name the missing affordance and the smallest public evidence that would change the verdict. Implementation discussion can use mioroute issue #2.
Stop rule
If the interface only looks friendly to a human but hides prompts, status, permissions, errors, readback, or rollback, keep it in observe/draft. Do not praise it as “agent-ready” until one task can pass through the structure.