"We cannot see ourselves aright until we see ourselves as nature's very own cyborgs." (Andy Clark, Natural-Born Cyborgs)
That line is a description of where we already are.
The next step is humans and machines working as one system. The merge is already here, and it is accelerating.
The most surprising thing about better AI
The better AI gets, the more I prompt.
Not because it's failing. Because it expands what I can attempt in a day. When the tool becomes more capable, the scope of the work expands with it. You stop using AI to help with tasks and start using it to run several workstreams at once.
Prompting becomes less like typing. More like management.
This is why "full autonomy" is a confusing destination. The real direction is humans doing more through machines, not humans who never prompt again.
"Full autonomy" is a stack
In practice, autonomy depends on a stack of layers underneath: the software (prompt, interface), the hardware (compute, networks, sensors, power), and the operations that keep it running (updates, governance, supply chain).
Break a link and the system becomes brittle. Most "autonomous" systems are really dependent autonomy: they look self-sufficient until one layer underneath them fails.
This is why Apple's local-first architecture matters beyond privacy marketing. On-device processing is the default, with Private Cloud Compute as the extension for heavier requests, and the design keeps personal data inaccessible even to Apple. That's an autonomy design decision. It defines where your intelligence can live, and who can touch it.
Safety is the interface
Why does safety talk feel discouraging? Because it sounds like bureaucracy that slows you down.
The merge gets safer by designing the exact interface between machine action and human judgment, not by writing policies.
In these systems, the key moment is escalation, when the machine hits uncertainty, consequences spike, or something turns irreversible.
That handoff is where safety lives: the line separating a draft from a sent message, a recommendation from a payment, a simulation from a moving vehicle.
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 dangerous failure mode is AI doing irreversible things without a real checkpoint.
What the merge looks like in the physical world
Phantom Auto was built around an honest idea: machines can operate until uncertainty shows up, and then humans take over.
The platform gave remote workers real-time 360-degree visibility around a vehicle, letting them jump between vehicles with the click of a button. That is the merge in one frame: a human nervous system, judgment, context, responsibility, connected to a machine body through a low-latency interface.
Phantom Auto shut down after failing to secure new funding. But 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. The pattern Phantom Auto pioneered is now visible across the AV industry.
The need for safe handoffs never disappears. It just moves to wherever the world is forced to learn it again.
The same pattern, now in knowledge work
At mixus, people set up agents through chat or email. They can run single-step or multi-step workflows from the inbox. The design is deliberately "colleague-in-the-loop": humans stay in control on mission-critical work as a safeguard against agent failure.
Machines do most of the work; humans own the line where consequences become real.
The merge will move fast. Just not at one speed.
Open and unchecked systems can move fast because they ignore boundaries. OpenClaw, an open-source AI agent framework, crossed 145,000 GitHub stars within weeks of launch. Soon after, security researchers found a one-click remote code execution flaw, plaintext credential storage, and hundreds of malicious skills in its marketplace. Unsafe speed can win early users. But in high-consequence industries, that trade-off is unacceptable.
The speed depends on the cost of being wrong.
Fast speed: where mistakes are cheap and reversible. Drafts and plans, open-ended exploration, internal or simulated runs.
Careful speed: where mistakes are expensive and irreversible. Health and identity, money and legal commitments, external communication and physical motion.
Knowing which speed you're operating at is the skill. Most failures happen when someone treats a careful-speed decision like a fast-speed one.
What we're actually becoming
When you prompt an agent, it can draft a contract, research a market, scaffold an architecture, and send back finished work. You directed a system doing what used to require a team. That loop is getting shorter and wider every month.
Andy Clark was right. We have always been cyborgs. We offloaded memory to cave walls, navigation to compasses, arithmetic to calculators. Every generation absorbed its tools so completely that the previous version of "human" became unrecognizable.
The machine extends you: your judgment moves through software you could never run alone.
What can machines execute without asking? What requires human authorization? Who controls the intelligence layer underneath? Those are the questions that decide whether the merge is safe enough to use.