---
author: "Mukul Chugh"
canonical: "https://mukulchugh.com/blog/human-in-the-loop-agent-plans"
description: "I want agents that make review cheap, not agents that skip the person who owns the blast radius."
image: "https://mukulchugh.com/og/blog/human-in-the-loop-agent-plans.png"
language: "en"
published: "2025-07-19"
title: "Agents draft. Humans release."
type: "article"
---

# Agents draft. Humans release.

Most demos end the same way: the agent writes the plan, merges the PR, sends the email. Then you remember production has irreversible actions and people who own the fallout.

My default is boring: agents **propose**, people **release**.

That means the unit of work isn't "chat output." It's a reviewable proposal: what changed, what evidence was used, what accepting it actually does. Review should feel closer to code review than dismissing a notification.

Automate gathering context and drafting options. Keep humans on sends, external communication, and anything that spends reputation. Speed without a gate is just faster mistakes.

---

[View this page](https://mukulchugh.com/blog/human-in-the-loop-agent-plans) · [Site index](https://mukulchugh.com/llms.txt)