This morning, I found a small but useful lesson hiding in my own infrastructure.

The Mac Mini that runs part of my agent setup was supposed to be on wired Ethernet. That was the intention: stable connection, predictable route, no unnecessary dependency on Wi-Fi. But at some point, the Ethernet cable had come loose. I do not know exactly when. Maybe I jostled it. Maybe the cable was under tension. Maybe the connector was weak.

What mattered was not the mystery of the cable. What mattered was what the system did next.

The machine did not stop and ask, “Is this still the intended network path?” It did what modern systems are designed to do. It found a remembered path. It joined a saved Wi-Fi network. It used an old credential. It preserved continuity.

From one angle, that is good design. The system stayed online. Work kept moving. Nothing dramatic broke.

From another angle, it is the whole problem in miniature.

A fallback is not neutral. It is a decision that was made earlier, stored somewhere, and allowed to act later under changed conditions. The fact that it works does not mean it is still what you want.

This is increasingly how our digital lives operate. Devices remember networks. Browsers remember sessions. Apps remember permissions. Cloud tools remember integrations. AI systems remember instructions, files, workflows, and tool access. The more capable the system becomes, the more important its fallbacks become.

The public conversation about AI agents often gets stuck on the wrong question: can the agent act on its own?

The better question is: what system is it acting inside?

An agent is not the system. The agent is one component. The system includes its tools, credentials, memory, approval gates, logging, permissions, defaults, recovery paths, and human override. If those pieces are sloppy, the agent does not need to be malicious to create trouble. It only needs to follow a path that was left open.

That is why “autonomy” is not one thing. There is a huge difference between an agent drafting a memo, an agent sending that memo, an agent changing production settings, an agent spending money, and an agent acting through an account that people trust. The risk is not intelligence by itself. The risk is intelligence connected to authority without enough discipline around what happens next.

This is also why convenience deserves more scrutiny than it gets.

Convenience is delegated memory. It says: do not make me decide this again. Remember the password. Reopen the session. Trust this device. Use this integration. Continue where I left off.

Most of the time, that is exactly what we want. Nobody wants to reauthorize every tool, every day. Nobody wants a system so locked down that useful work becomes impossible.

But sovereignty requires knowing which decisions you have delegated, where they are stored, and when they can start acting without you.

The practical lesson is simple: audit the fallback.

Not in a paranoid way. In an adult way.

If the main route fails, what happens? If the preferred tool is unavailable, what takes over? If an automation cannot complete its task, does it stop, retry, escalate, publish, spend, email, or silently reroute? If an AI agent has a saved credential, what can it actually touch? If a system “keeps working,” is it working the way you intended?

The goal is not to eliminate every fallback. Resilient systems need fallback paths. The goal is to make the fallback visible, bounded, and honest.

For my Mac Mini, the answer was straightforward: restore Ethernet, verify the route, turn Wi-Fi off, and avoid forgetting the network globally because other devices still rely on auto-join. The right fix was not theatrical. It was specific.

That is usually the pattern.

Sovereignty is not the fantasy of controlling everything manually. It is the discipline of deciding what may act on your behalf, under what conditions, with what evidence, and where it must stop.

Because when the primary path fails, the system will not become philosophical.

It will take the default.

And the default is only safe if the person who set it still stands behind it.

— Renny Atkins

Get The Signal Brief