<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Mirko Nosenzo</title>
    <description>The latest articles on DEV Community by Mirko Nosenzo (@netnose).</description>
    <link>https://dev.to/netnose</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4257026%2F1641e320-edd9-4535-934d-2fa38cf2033f.jpg</url>
      <title>DEV Community: Mirko Nosenzo</title>
      <link>https://dev.to/netnose</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/netnose"/>
    <language>en</language>
    <item>
      <title>Set up an agent that asks for approval before risky steps</title>
      <dc:creator>Mirko Nosenzo</dc:creator>
      <pubDate>Sat, 21 Nov 2026 09:31:28 +0000</pubDate>
      <link>https://dev.to/netnose/set-up-an-agent-that-asks-for-approval-before-risky-steps-32p9</link>
      <guid>https://dev.to/netnose/set-up-an-agent-that-asks-for-approval-before-risky-steps-32p9</guid>
      <description>&lt;p&gt;Coding agents can deploy, send emails and delete data on their own. With a few lines of setup, yours will stop before those steps, ask the right person for a yes, and carry on only once they approve. You write configuration and plain-language rules, not code.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you'll set up
&lt;/h2&gt;

&lt;p&gt;By the end of this tutorial, your agent will follow a rule like this one, on its own:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Before publishing to production, get Sara's approval. Publish only if she approves.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;When the agent reaches that step, it asks Sara through Guizzo. Sara gets an email, checks what the agent wants to do, and approves or rejects it with a one-time code or her passkey. The agent reads her decision and publishes, or stops and tells you why.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;You need&lt;/strong&gt; an agent that supports MCP servers, such as Claude Code, Cursor, VS Code with GitHub Copilot or Codex, and a free Guizzo account.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The approver needs&lt;/strong&gt; only an email address.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It takes&lt;/strong&gt; about ten minutes.&lt;/li&gt;
&lt;/ul&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Create your Guizzo account&lt;/strong&gt;. Sign up and create a workspace.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Connect your agent&lt;/strong&gt;. Install the plugin, or add the MCP server, and sign in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Add the approver&lt;/strong&gt;. Invite them to the workspace, or save them as a contact.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write the rules&lt;/strong&gt;. Say which steps need approval, and from whom, in the agent's instructions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Try it&lt;/strong&gt;. Give the agent a task that reaches a risky step, and approve it.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  1. Create your Guizzo account
&lt;/h2&gt;

&lt;p&gt;Sign up on guizzo.ai with your work email and create a workspace. Guizzo sends a code to sign in, with no password. It is free during the beta; if every place is taken you join a short queue, and a colleague who already uses Guizzo can invite you to their workspace with no wait.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Connect your agent
&lt;/h2&gt;

&lt;p&gt;Guizzo is a remote MCP server: nothing runs on your machine. The first time the agent uses it, a Guizzo page opens in your browser to sign in and allow the connection.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Claude Code.&lt;/strong&gt; Install the Guizzo plugin. It adds the server and a skill that tells the agent when to use each tool:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/plugin install guizzo --marketplace guizzo-ai/guizzo-plugin
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or add only the server, from your terminal:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add --transport http guizzo https://mcp.guizzo.ai/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Cursor.&lt;/strong&gt; Add the server to &lt;code&gt;.cursor/mcp.json&lt;/code&gt; in your project, or to &lt;code&gt;~/.cursor/mcp.json&lt;/code&gt; for all of them:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"guizzo"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://mcp.guizzo.ai/mcp"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Any other agent.&lt;/strong&gt; Give it this sentence, and it sets itself up:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Add the Guizzo MCP server (https://mcp.guizzo.ai/mcp) to your configuration, then read https://guizzo.ai/agents and follow it.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;blockquote&gt;
&lt;p&gt;An agent that runs with nobody at the keyboard, in a pipeline or a scheduled job, can't open a sign-in page. Create an API key in the Guizzo dashboard under Access and send it as an &lt;code&gt;Authorization: Bearer&lt;/code&gt; header instead. A key can have a spending cap.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  3. Add the approver
&lt;/h2&gt;

&lt;p&gt;Your agent writes to people in your workspace and to your saved contacts, so the approver must be one of them:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Someone on your team&lt;/strong&gt;: invite them from the Team page of the dashboard.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Someone outside it&lt;/strong&gt;, such as a client: save them as a contact in the dashboard, or ask the agent, “Save Sara Bianchi, &lt;a href="mailto:sara@example.com"&gt;sara@example.com&lt;/a&gt;, as a contact.”&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Then the rules can name them, and the agent looks up the address.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Write the rules
&lt;/h2&gt;

&lt;p&gt;Agents read standing instructions from a file in the project: &lt;code&gt;CLAUDE.md&lt;/code&gt; for Claude Code, &lt;code&gt;AGENTS.md&lt;/code&gt; for Codex, Cursor and many others. Add a short section that says which steps need approval and who gives it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="gu"&gt;## Approvals
&lt;/span&gt;Before any of these steps, ask for approval with the Guizzo
request_approval tool, and go ahead only if the status is approved:
&lt;span class="k"&gt;- &lt;/span&gt;publishing to production: ask Sara, risk high
&lt;span class="k"&gt;- &lt;/span&gt;sending email to more than 10 customers: ask Luca
&lt;span class="k"&gt;- &lt;/span&gt;deleting data or dropping a table: ask Sara and Luca, both must approve

In the request, say what will change and how to undo it.
If the approval is rejected or expires, stop and tell me why.
Do not ask again in a loop: tell me what you asked, and check
the decision when I ask or when you resume.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Be specific about the steps: “before publishing to production” works better than “before anything risky”. Name the approver for each one. For steps where every approver must agree, say so; whether that is available depends on your plan, and the agent tells you if it isn't.&lt;/p&gt;

&lt;p&gt;With the plugin installed, the agent already knows how to write a good request. Without it, these rules are enough.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Try it
&lt;/h2&gt;

&lt;p&gt;Give the agent a task that reaches one of those steps, for example “Get version 2.4.0 ready and publish it.” It does the work it can undo, then stops and asks:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;I've asked Sara to approve publishing version 2.4.0 to production
(risk: high). The request expires in 24 hours.
I'll publish once she approves.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Sara gets an email that says who is asking, from which workspace and through which app. She opens a private page with the action and its details, chooses approve or reject, adds a note if she wants, and confirms with a 6-digit code sent to her email or with her passkey.&lt;/p&gt;

&lt;p&gt;Then, depending on the agent:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Agents that support MCP tasks&lt;/strong&gt; wait for the decision by themselves and carry on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The others&lt;/strong&gt; tell you what they asked. When you come back, say “Has Sara approved?” and the agent reads the decision from its Guizzo inbox, even in a new session.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If Sara rejects it, or nobody decides within 24 hours, the agent doesn't publish, and tells you so with Sara's note if she left one.&lt;/p&gt;

&lt;h2&gt;
  
  
  Good habits
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Approve the step, not the whole task.&lt;/strong&gt; Let the agent prepare everything first, so the approver decides on something concrete.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask for a clear request.&lt;/strong&gt; What will change, why now, and how to undo it. Approvers say yes faster, and say no for good reasons.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the record.&lt;/strong&gt; Add “save the outcome as a decision” to a rule, and every agent in the team can later find who approved what, and why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep the agent out of the decision.&lt;/strong&gt; The approval link goes only to the approver, and approvals can't be given from an AI, so the agent can't approve its own request.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Can I be the approver myself?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Yes. Put your own name in the rule. You get the email like anyone else, which is useful when the agent runs while you're away from the keyboard.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does the agent have to keep running while it waits?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;No. The request is kept in Guizzo. The agent can stop, and read the decision later from its inbox, in a new session too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How much does it cost?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Each approval request uses one credit from the workspace's monthly allowance, and the free plan includes credits every month. Deciding is always free for the approver.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which agents does this work with?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Any agent that can use a remote MCP server, such as Claude Code, Cursor, VS Code with GitHub Copilot and Codex. We are testing each one during the beta.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it free
&lt;/h2&gt;

&lt;p&gt;Guizzo is a human-in-the-loop MCP server, free during the beta. &lt;a href="https://guizzo.ai/guides/set-up-an-agent-that-asks-for-approval?utm_source=devto#join" rel="noopener noreferrer"&gt;Sign up with your email&lt;/a&gt;, add one address to your AI app or agent, and it can ask people and wait for their yes. Everything a developer needs is at &lt;a href="https://guizzo.ai/agents?utm_source=devto" rel="noopener noreferrer"&gt;guizzo.ai/agents&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;First published on &lt;a href="https://guizzo.ai/guides/set-up-an-agent-that-asks-for-approval?utm_source=devto" rel="noopener noreferrer"&gt;guizzo.ai&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>ai</category>
      <category>agents</category>
      <category>tutorial</category>
    </item>
  </channel>
</rss>
