Make a background agent wait for a person's approval.

An agent that runs on its own, in a pipeline, a scheduled job or a long coding session, has nobody watching the chat when it reaches a risky step. With Guizzo it asks the right person, waits as long as it takes and only carries on once that person has said yes.

Get started Read the story

Release agentClaude Code· running on its own

request_approval(
  approvers: "sara@acme.example",
  action:    "Publish v2.4.0 to production",
  risk:      "high",
  expires_in_hours: 2
)

SaraPhone· 9 min later

Approvedwith a one-time code

"Go ahead, the 18:00 backup is done."

The agent picked up where it stopped and shipped.

The story

From the first question to the last step.

1
The agent reaches a risky step
A release agent in Claude Code has version 2.4.0 ready: tests pass and the notes are written. Publishing to production is the one step it may not take alone.
2
It asks Sara for a sign-off
The agent asks the tech lead, Sara, to approve the release. It says what will happen and how risky it is.
3
Sara decides from her phone
Sara gets an email, or a notification if she added Guizzo to her home screen. She opens a private page that shows the release, and approves with a one-time code or her passkey.
4
The agent picks up the decision
Clients that support MCP tasks wait for the decision by themselves. Others check the agent's inbox later. Either way, the agent never has to keep a chat open.
5
It ships, or it stops
A yes lets the agent publish. A no, or no answer before the request expires, stops it, and the agent reports that nobody approved.

What makes it work

What Guizzo does here.

Proof at the moment of decision

A fresh code or a passkey shows the person deciding is the one the agent asked. A forwarded link can't approve.

Approvals

The agent can't approve itself

Approvals are given only on their own web page, never by an AI. The private link goes to Sara, never back to the agent.

Security

Every decision is kept

The agent gets who decided, how they proved it and their note. It can also record the decision for the rest of the team.

Team memory

In practice

Say it to your agent.

Once Guizzo is connected, your AI knows when to use it. You don't need to name a tool.

“Before you publish to production, get Sara's approval.”
“Don't delete anything until both Luca and Anna have approved.”
“Ask finance to approve any payment over 500 euros before you send it.”

Questions

Good to know.

How does an agent with no chat window connect?

Agents in a client such as Claude Code or Cursor connect like any AI app. Scripts and agents that run on their own use an API key from the dashboard, and the key can have a spending cap.

Does the agent have to stay running while it waits?

No. The request lives in Guizzo, not in the agent. An agent can stop, start again later and read the decision from its inbox.

What are MCP tasks?

An extension of the Model Context Protocol for work that finishes later. When a client supports it, an approval request becomes a task that completes when the person decides, so the client knows when to carry on.

Can several people have to agree?

Yes. The agent can go ahead on the first yes, or wait until every approver has agreed.

Keep exploring

The features in this story.

Free beta

Give your agent someone to ask.

Sign up with your email and create your workspace. We open Guizzo to new teams a few at a time: if every place is taken, you join the queue and we write when yours is ready.

Make a background agent wait for a person's approval.

From the first question to the last step.

The agent reaches a risky step

It asks Sara for a sign-off

Sara decides from her phone

The agent picks up the decision

It ships, or it stops