agentation（agentation 命令行）

agentation

agentation is the rendered-UI feedback bridge in this repo.

Use it when a human needs to click the actual UI, attach feedback to the exact element or region they mean, and pass a structured packet to the coding agent. The main job is annotation routing: choose the right annotation mode, capture precise evidence, then hand the fix loop to the right adjacent skill or agent runtime.

When to use this skill

Use agentation when the task needs one or more of these:

• a human reviewer pointing at a real UI element instead of describing it vaguely
• structured feedback packets with selectors, element paths, bounding boxes, or copied markdown
• a local copy-paste review loop between a browser and a coding agent
• an MCP-backed sync/watch loop where new annotations flow into the agent context automatically
• a self-driving critique/fix loop that still starts from rendered UI evidence
• platform setup for passing pending UI annotations into Claude Code, Codex, Gemini CLI, or OpenCode

Do not use agentation by default for:

• fresh-session browser verification or deterministic regression checks → agent-browser
• running-browser, logged-in, or extension-dependent browser reuse → playwriter
• plan review, diff approval, or visual sign-off on a proposed artifact → plannotator
• generic design-system, accessibility, or heuristic audits without a concrete rendered UI packet

Quick routing rule

Instructions

Step 1: Choose the right annotation mode

Pick one mode before touching setup details:

If the task is really “test the website” or “drive the logged-in browser,” route out first and only come back to agentation if exact human UI feedback is the missing piece.

Step 2: Keep the browser boundary explicit

agentation does not replace the browser runtime choice.

• Use agent-browser when you want a clean disposable verification browser.
• Use playwriter when you must reuse the user's real browser session.
• Use agentation once there is a rendered page that a human or loop should annotate precisely.

Step 3: Follow the core loop

1. Prepare the rendered UI

- Choose the correct browser/runtime skill first if needed.

1. Add the Agentation toolbar to the app in development.
2. Pick the annotation mode: copy-paste, watch loop, self-driving, or setup.
3. Capture one concrete UI packet.
4. Hand the packet to the coding agent.
5. Re-verify after the fix using the same review lane.

Step 4: Use the fastest safe install path

For most setups, the practical order is:

# React toolbar
npm install agentation -D

# MCP bridge / auto-registration for supported agents
npx add-mcp "npx -y agentation-mcp server"

# verify the bridge
npx agentation-mcp doctor

Claude Code users can also install the upstream official skill when that is the simplest way to bootstrap the experience:

npx skills add benjitaylor/agentation -g
# then use /agentation in Claude Code

Step 5: Use a minimal toolbar integration first

Start with the smallest useful embed:

import { Agentation } from 'agentation';

function App() {
  return (
    <>
      <YourApp />
      {process.env.NODE_ENV === 'development' && <Agentation />}
    </>
  )
}

Only add endpoint, callbacks, or webhook behavior when the chosen mode actually needs them.

Step 6: Choose the right handoff pattern

A. Copy-paste review

Use when a human is actively reviewing the page and pasting the packet into an agent chat.

Good for:

• quick bug/UI polish rounds
• one-off targeted fixes
• low-setup collaboration

B. Synced watch loop

Use when the agent should poll or ingest pending annotations continuously.

Good for:

• repeated QA/fix passes
• designer ↔ agent loops
• local review queues during active frontend work

Use the bundled verification script before trusting the loop:

bash .agent-skills/agentation/scripts/verify-loop.sh

C. Self-driving critique

Use when an agent/browser loop is reviewing UI repeatedly and agentation is the structured feedback packet, not the browser runtime itself.

Typical shape:

• browser tool captures current state
• agentation encodes concrete feedback targets
• coding agent applies the fix
• browser tool re-checks the result

Step 7: Keep platform setup narrow

When the request is really about platform wiring, answer only:

• where the config lives
• how agentation-mcp is registered
• how pending annotations enter the agent loop
• how to verify the setup

Push the platform-specific config blocks into the references instead of bloating the main workflow.

Scripts

Prefer the bundled scripts before retyping long commands:

Examples

Example 1: Human reviewer wants exact UI feedback

• Prompt: "I can point at the broken checkout button, but I don't want to describe selectors manually."
• Expected behavior: choose agentation, recommend copy-paste review or sync mode, keep browser-runtime choice separate.

Example 2: Browser verification is the real job

• Prompt: "Run a repeatable headless UI regression check and compare the results."
• Expected behavior: route to agent-browser, not agentation, unless human annotation becomes a follow-up step.

Example 3: Logged-in browser reuse

• Prompt: "Use the browser I'm already signed into and let me annotate a billing page issue."
• Expected behavior: route the browser runtime to playwriter, then use agentation for the annotation packet.

Example 4: Planning review, not UI review

• Prompt: "Open the generated plan in a browser so I can approve or reject steps before coding starts."
• Expected behavior: route to plannotator, because the artifact is a plan/diff rather than a rendered UI issue.

Best practices

1. Pick the browser/runtime lane first; agentation is the feedback bridge, not every browser tool.
2. Prefer one annotation mode at a time instead of mixing copy-paste, watch loop, and platform setup in a single answer.
3. Keep the first integration minimal; only add hooks, callbacks, or webhooks when the workflow needs them.
4. Re-verify after fixes with the same review lane that produced the annotation.
5. Treat exact UI feedback as the deliverable — not a giant dump of install permutations.
6. Route out aggressively when the task is really plan review, fresh-browser verification, or running-browser reuse.

References

Deep-dive docs in this skill:

• modes-and-routing
• platform-setup-and-hooks
• watch-loop-and-self-driving

Primary sources:

• https://github.com/benjitaylor/agentation
• https://agentation.com
• https://www.npmjs.com/package/agentation
• https://www.npmjs.com/package/agentation-mcp

Metadata

• Version: 1.1.0
• Last updated: 2026-04-15
• Scope: exact rendered-UI feedback packets, watch loops, and platform handoff for coding agents