Human in the loop won't last

I think human in the loop is a phase, and the companies that get through it first will outrun everyone else.

Today almost every agent action starts with a person at the keyboard, and when the agent wants to write something, a person clicks approve. Our team was on a call with a large investment firm that put it well. Read is solved for them. Write is where it breaks, and a person approving every write doesn't fix it, because people approve everything. "Even if they're really good employees, they're just going to say yes to everything eventually."

Approval prompts are the new cookie banners

You read the first few. By the tenth you click yes to make it go away, and the log still says a person approved it. The prompt shows a tool name and some JSON, and the person clicking has no idea what the agent did three steps earlier.

The loop also costs you the reason you wanted an agent. An agent that waits for my approval moves at my speed. Ask anyone who's waited on me for a follow-up.

Agents hand work to agents

And the loop is about to lose its person. Agents are starting to hand work to other agents. Say a triage agent hands a ticket to a coding agent, which calls a deploy agent. The person who opened the ticket is gone by the second hop, which runs as a service account with nobody's identity on it. There's no one to show the prompt to. Service accounts are the wall every security team hits.

Every hand-off still needs the three answers a person used to give by clicking yes: who is this agent, is it allowed, and did it do what it said? Next, the system has to give them, on every hand-off, in milliseconds.

Billions to quadrillions

Once nobody's at the keyboard, traffic stops growing at human speed. A lot of what we do slowly through websites and UIs moves to agents talking to each other, inside the company first, then across companies. Billions of agent events a day become trillions, then quadrillions. At that scale, latency and reliability become core product for anything that sits in the path.

No team approves a quadrillion of anything.

Nobody can click approve that fast

Take two companies with the same models. One keeps a person on every write. The other lets its agents decide, checks every step against what the person asked for, and every decision lands five seconds faster. Five seconds sounds like nothing until it's on every step of every session, across every agent in the company. That compounds into ten times faster, then a hundred. I don't think the first one keeps up.

People still set the intent and the limits, and they get pulled in when an agent drifts from them. Everything else runs at machine speed. That check is what we build at Runlayer.

I want to run a company where agents hand work to each other all day, and a person only gets pulled in when it matters.