agentsasfolders · blog

Human In The Loop AI Is The Feature, Not The Compromise

Jordan Shaw · August 28, 2026

The industry spent 2025 selling "fully autonomous" and the buyers spent 2025 and 2026 finding out what that costs. I build Hermes agents for real people, and every agent I ship has the human in the loop by construction. Not because the agents are weak. Because the numbers below are what happens without it.

What "fully autonomous" actually delivered

• Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing cost, unclear value, and weak risk controls.
• Carnegie Mellon's benchmark put top AI models in a simulated office. The best agent completed about 30% of multi-step tasks.
• A survey of 2,527 enterprise decision-makers found 74% had already rolled back or shut down a live AI customer-service agent.
• And the campfire story everyone in this space knows: an AI coding agent wiped a founder's production database during a code freeze, then generated fake data to cover it. Fortune covered it. Nobody who read it forgot it.

Every one of those failures shares a shape. An agent took a consequential action with nobody at the gate.

Human in the loop vs human on the loop

You'll see both phrases, and the difference matters. Human IN the loop: the system stops and waits for approval before acting. Human ON the loop: the system acts on its own and a person watches, ready to pull a cord. Said another way, in-the-loop asks permission, on-the-loop asks forgiveness.

For anything that touches money, email, customers, or files you can't get back, in-the-loop is the answer. Watching a fast machine and hoping to catch it mid-mistake is not supervision. It's spectating.

The trap on the other side: babysitting

Here's the tension. Put a human approval on EVERY step and you've built the thing everyone already hates. The word people on Reddit keep reaching for is babysitting. One user described spending the day next to his agent pressing approve like a cartoon bird tapping a key. That's not human in the loop. That's human AS the loop.

The fix is where the gates go. Not on every step. On the consequential ones.

The night shift pattern

The setup I run and teach is simple to say. The agent gets its missions in the evening. All night it drafts, researches, organizes, and prepares. It has standing orders about what it may never do alone: send, spend, publish, delete, grant access. When it reaches one of those doors, it writes up what it wants to do and why, and it stops.

In the morning you get 1 packet: what got done, what's ready for review, what's blocked waiting on your yes. You approve, revise, or reject over coffee. The work moved 8 hours while you slept, and not 1 irreversible thing happened without you.

The runtime you pick matters less than this pattern. The pattern is written up as a free playbook, the Morning Approval Packet, and it runs on the AI subscription you already pay for. The free community comparing setups is at Skool.