At Phantom Auto, a remote operator in California could drive a vehicle in Las Vegas. The machine handled routine motion. A person stepped in when the situation became uncertain. We built the company around that handoff.
AI agents are bringing the same arrangement to knowledge work.
"We cannot see ourselves aright until we see ourselves as nature's very own cyborgs." (Andy Clark, Natural-Born Cyborgs)
Clark wrote that about tools in general. Agents make it literal: a person and a machine working as one system, with a line between them that someone has to draw.
What I ask of the tools
The better AI gets, the more I prompt.
Better models make me more ambitious. I give them more work, then spend more time deciding what to do with the results.
Yet someone still has to run the servers, maintain the software, and deal with failures, and I don't see that work disappearing.
The handoff
Which decisions should the system bring back to a person, and what does that person need to see? The merge gets safer when the handoff is explicit: the machine hits uncertainty, consequences spike, or something turns irreversible, and the work crosses to the person with everything the machine saw. That line separates a draft from a sent message, a recommendation from a payment, a simulation from a moving vehicle.
At Phantom Auto, the platform gave remote workers real-time 360-degree visibility around a vehicle and let them move between vehicles with the click of a button. They needed to see enough of the situation to take control when asked.
Phantom Auto shut down after failing to secure new funding. The pattern didn't die with the company. Today, Waymo and Tesla each employ hundreds of remote human operators who assist their autonomous vehicles in real time. As machines improve, the handoff moves to harder decisions.
At mixus, an agent can prepare work in the email thread, then wait for the person who needs to approve it.
A lot of modern AI failures come from crossing that line without realizing it. In April 2025, Cursor's AI support bot invented a policy, "one device per subscription," and users started canceling. The words acted like an official commitment, even though no human ever made it.
The cost of a mistake
I can let an agent try several drafts or explore a plan; while the work stays internal, the mistakes are drafts nobody has read.
Sending a message to a client, making a payment, or moving a vehicle needs more care. Other people are affected, and undoing the action may not undo the harm.
I use these tools every day. I still have to decide which decisions I am comfortable handing over.
The machine extends your reach. The hard part is deciding where it must stop and ask.