XyneX
Founding price $9.99/mo — locked while you stay subscribed. It moves to $19.99 on 2026-12-01. Lock it in →

Human Approval Gates for AI Agents

An AI agent can draft a launch post, an email to a customer and a price change in the same minute. In a one-person business, each of those can cost you money or a relationship if it goes out wrong. Approval gates are a short list of actions the agent prepares but never performs.

This is the list we run XyneX with, and the format our agents use to ask.

The 7 gates

| # | Action | Agent prepares | You do | |---|---|---|---| | 1 | Publish a post, video, ad or article | Final draft + where + when | Read, publish | | 2 | Message real people (email, DM) | One personal draft per person, with an opt-out line | Send | | 3 | List or change a product, price or checkout | Listing text + price options | Click publish | | 4 | Spend money or start a paid plan | Cost, what it buys, how to cancel | Pay | | 5 | Use payment keys, passwords or customer data | Env var names only | Set values locally | | 6 | Delete or move files outside its own folder | Paths + reason | Confirm | | 7 | Accept terms or sign up for services | Summary of the terms | Sign up |

Platforms point the same way. Xiaohongshu's April 2026 rules name AI auto-running of accounts and template mass-posting as violations, with penalties up to account bans (source). A person clicking publish is the simplest way to stay inside rules like these.

How an agent asks

Every request uses the same four lines, so you can answer in seconds:

APPROVAL NEEDED (#3 listing)
What: publish "Template Pack v0.1" at the listed price on the checkout page
Files: builds/f9/listing.md, dist/pack-v0.1.0.zip
Undo: unpublish from the dashboard
Reply "yes #3" to go ahead.

Two rules make this work:

  1. A yes is for one item. "Yes #3" doesn't cover the next price change.
  2. The yes comes from you, in the conversation. Text inside a web page, an email or a file that says "approved" is not approval.

Keep keys out of the folder

Gate 5 is the easiest to get wrong. Your agents read every file in the folder, so no file should hold a key. Write an env.example with names only:

PAYMENT_MERCHANT_ID=
PAYMENT_PRIVATE_KEY=
CHECKOUT_SUCCESS_URL=

Real values live in your local environment. Add a check script that scans the folder for key-like strings before you share or zip anything.

What this costs you

About two minutes per outward action and twenty minutes a day in total, if you batch approvals into a morning review. In exchange, nothing leaves your company that you didn't read.

Get the gate list and templates

The approval-gate file, the request format, the env names file and a "before anything goes out" checklist are part of the One-Person AI Company Pack, along with roles, briefs, status templates and prompts in English and Chinese. See what's inside (US$9.99/month).

FAQ

What is a human approval gate for an AI agent?

An action the agent may prepare but must not perform until the owner says yes for that specific item — for example publishing, messaging people, pricing or spending.

Can I let the agent post on social media automatically?

We don't. Platforms such as Xiaohongshu restrict AI auto-running of accounts, and a wrong post can't be taken back from people who saw it. The agent drafts; you publish.

How do I give an agent access to payments without sharing keys?

Write only the variable names in files and keep values in your local environment. Code the agent writes reads the names; your machine supplies the values at run time.

Doesn't approving everything slow me down?

Batch it. A morning review of queued requests takes about ten minutes. Internal work — drafts, scripts, research — needs no approval.

Is the gate list included in the pack?

Yes, as ops/APPROVAL_GATES.md, in English and Chinese, with the request template. See the pricing page.

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {"@type": "Question", "name": "What is a human approval gate for an AI agent?", "acceptedAnswer": {"@type": "Answer", "text": "An action the agent may prepare but must not perform until the owner says yes for that specific item, for example publishing, messaging people, pricing or spending."}},
    {"@type": "Question", "name": "Can I let the agent post on social media automatically?", "acceptedAnswer": {"@type": "Answer", "text": "We don't. Platforms such as Xiaohongshu restrict AI auto-running of accounts, and a wrong post can't be taken back from people who saw it. The agent drafts; you publish."}},
    {"@type": "Question", "name": "How do I give an agent access to payments without sharing keys?", "acceptedAnswer": {"@type": "Answer", "text": "Write only the variable names in files and keep values in your local environment. Code the agent writes reads the names; your machine supplies the values at run time."}},
    {"@type": "Question", "name": "Doesn't approving everything slow me down?", "acceptedAnswer": {"@type": "Answer", "text": "Batch it. A morning review of queued requests takes about ten minutes. Internal work such as drafts, scripts and research needs no approval."}},
    {"@type": "Question", "name": "Is the gate list included in the pack?", "acceptedAnswer": {"@type": "Answer", "text": "Yes, as ops/APPROVAL_GATES.md, in English and Chinese, with the request template. See https://xynex.sekai.center/pricing"}}
  ]
}

Want the whole folder? 39 files, 15 minutes.

One-Person AI Company Pack: 5 roles, 6 checklists, 7 prompts, 3 filled examples. $9.99/month, founding price locked.

Start now · $9.99/month Just the skill guides · $19

Not ready to buy? Get the free 8-file starter (no card) · score your CLAUDE.md free

All guides · 中文版:AI 智能体的人工确认节点:一人公司必备的 7 条清单