# Agentic Experience Checklist

Use this checklist when designing, building, or reviewing systems where an AI agent is a first-class user.

## Entry Point

- Is there a canonical manifest or equivalent first read?
- Can a fresh agent understand the system without out-of-band guidance?
- Are human-facing and agent-facing surfaces intentionally separated when needed?
- Does the system avoid requiring screen-reading when structured state exists?

## Capabilities

- Are capabilities typed and discoverable?
- Are inputs, outputs, defaults, permissions, and side effects explicit?
- Are examples included for non-obvious operations?
- Is every operation named by what it does, not by the UI gesture that triggers it?

## State And Memory

- Can an agent resume without re-reading everything?
- Are freshness, version, and content hash exposed?
- Is important state available as data rather than pixels?
- Are context boundaries explicit?

## Safety And Recovery

- Are destructive or external side effects marked before execution?
- Are operations idempotent where possible?
- Do errors include class, cause, retryability, and suggested next action?
- Can an agent recover without guessing?

## Output

- Can an agent cite sources or endpoint paths?
- Can a human distinguish facts from assessment?
- Are uncertainty and missing data represented directly?