On a 25,000V overhead line, nobody is expected to spot a fault and react in time. Protection is designed in: a device sees the fault and opens the circuit, whether or not anyone is watching. Safety-critical industries made dangerous systems safe that way, with architecture rather than attention.
AI agents now write and ship code, move money and change infrastructure. Asking a person to watch every action does not scale, and it was never how those industries got safe. AI needs its own breakers.
Equipment first, people later
Circuit breakers first protected equipment. Devices that protect people came much later. That order is worth remembering, because the first question in any protection scheme is not how to stop the fault but what is being protected. A breaker sized to save a cable will not save a person.
AI is at a similar point. It is easy to build controls that protect the system itself. It is harder, and more important, to protect the user, an organisation’s data and the public.
The breaker is the easy part
A circuit breaker is a simple device. The hard part is the settings, and three questions decide them.
- What is being protected?
- The system, the user, an organisation’s data or the public. Each needs its own protection, and one agent may need several at once.
- What is the fault current?
- An electrical breaker measures current. An AI breaker needs a signal that rises when an agent is going wrong, early enough to act before the harm is done. Choosing that signal is most of the design.
- Who sets the trip points?
- Regulators, organisations and users need different things. The settings have to say whose rules apply, and which wins when they disagree.
Identify, isolate, remedy, then reinstate or lock out
When a breaker trips, the work is not over. The cycle is the same in any protection scheme: identify the fault, isolate the agent, remedy it, then reinstate it once it is proven safe, or lock it out.
Lockout matters as much as the trip. An auto-recloser tries to restore power after a fault, because many faults clear on their own. If the fault persists, it stops trying and stays open until a person intervenes. An agent that keeps failing the same check should be treated the same way: not retried forever, but stopped until someone looks.
Different systems, different rules
Not every AI system needs the same protection. A chat assistant, a coding agent and an agent that moves money or changes infrastructure carry different risks, in the same way that a bathroom, a factory and a transmission line follow different rules. AI will likely need its own classes of protection, each with settings that suit what it can harm.
How I work with it now
I run an agentic studio where AI agents are the engineering, design and QA team, and I run it on these principles. No agent signs off its own work: the builder never certifies itself. Work moves through stage gates with checks that must pass. Every claim is graded by its evidence before it is trusted. Production database changes need a human’s sign‑off, and anything touching customer funds gets two reviews.
Two more come straight from protection engineering. A fault that persists should stop the system until a person intervenes. And a new gate should observe before it enforces: it logs what it would have blocked before it blocks anything, so its settings are proven on real work first.
None of this is new. It is what protection engineering already knows, applied to a new kind of system. The open question is the settings, and that is the work I want to do.
