2 min read

Why We Added a Human Approval Step to Our Autonomous Telegram Agent

Why We Added a Human Approval Step to Our Autonomous Telegram Agent

What Happened

Last month our Hermes agent sent a sales follow-up to a customer who had already asked to be removed from our list.

How We Use Hermes

We run Hermes on Docker Compose and use it primarily in Telegram for support and sales follow-up.

The agent reads incoming messages, drafts replies, and sometimes sends them automatically. For most routine questions this worked well.

The Problem With Full Autonomy

But the agent doesn't always get context right. It missed a "do not contact" note buried in a thread. The message went out.

The customer complained. We spent a day on damage control. The agent had done exactly what we told it to do, but we hadn't given it a way to ask for help.

Why This Matters

For any business, an agent that sends wrong messages or performs wrong actions costs money and reputation.

In support, a wrong answer can escalate a ticket. In sales, a wrong follow-up can kill a deal.

The cost of doing nothing is not just errors. It's the risk that you can't let the agent do anything useful without constant supervision.

Autonomy without approval is just risk with a nice interface.

Adding an Approval Step

I added an approval step for any action that leaves the system: sending a message, updating a CRM record, or scheduling a meeting.

The agent now drafts the action and posts it to a Telegram approval queue. I tap Approve or Reject. The agent only executes after approval.

To keep it from slowing down the workflow, I set a few rules:

  • ✅ Approvals only for external actions. Internal lookups and reads are still fully autonomous.
  • 💡 Inline buttons in Telegram so approval takes one tap.
  • ⏱️ A timeout: if no one approves within 5 minutes, the action is dropped and logged.

These rules kept the approval step from becoming a bottleneck.

The Result

We've had zero wrong messages in two months. Response time stayed under 10 minutes because approvals are fast.

The agent still saves us time on drafting and data lookup. We reduced manual reply writing by about 60 percent.

Most importantly, I can let the agent act without watching every move. It asks when it needs to, and I approve in seconds.

AgentForge

AI agents for real work. If you want to see how this works in practice, visit our product site and blog.

Share
Serhii

Serhii