AI agents are no longer chat windows waiting for a question. They read tickets, create files, call interfaces, start builds and often work across multiple systems in the process. It is precisely this autonomy that makes them useful – and precisely this that raises the question of how a company can halt such a process again if in doubt.
The security provider Crowdstrike has announced a tool intended as an emergency stop for AI agents in the corporate network. Administrators should be able to stop running agents and prevent them from simply starting again. Technically this is a small building block, but in substance it is a clear signal: autonomy needs a counterforce.
Why a stop button is more difficult than it sounds
With a classic application the matter is simple: terminate the process, deactivate the service, done. With agents the case is different. An agent is usually not a single process, but an interplay of a language model, a workflow control, access credentials and a series of tools – that is, interfaces through which the agent actually changes something in other systems.
This gives rise to several practical problems. First, an agent often does not run where it has its effect: the control can lie in a cloud environment while the effects hit internal systems. Second, many agents are built in such a way that they restart themselves after an abort – via schedulers, watchdogs or queues. A terminated run is then not a stopped agent, but a briefly interrupted one. This is precisely the point at which the idea of actively blocking the restart comes in.
Third, an aborted agent leaves behind a half-finished state: a created data record without a corresponding entry, an opened pull request without a description, a migration that has begun. An emergency stop ends the activity, it does not undo it.
Four levels of control
It makes sense not to think of control as a single switch, but as a graduated chain. In projects, four levels can be distinguished:
- Identity: Every agent needs its own technical identity, not the access of an employee or a shared collective account. Only in this way can its actions be traced and withdrawn individually at all.
- Authorization: Rights are tailored to what the specific use case needs. An agent that pre-records invoices does not need write access to the user directory.
- Observability: Every tool call is logged – who, when, with which parameters, with which result. Without this trail, neither troubleshooting nor a clean follow-up is possible.
- Shutdown: A defined way to stop the agent, and a second one to prevent it from restarting. Both must work even when the person who set up the agent is currently unreachable.
The first three levels are classic topics of IT security. What is new is above all that they must now be applied to actors that make decisions independently and whose behaviour is not fully predictable.
Who may pull the plug?
An emergency stop is only as good as the organization behind it. In practice, three questions arise that should be clarified before the first productive agent.
Who decides? Named persons are needed who may switch off an agent without asking – and who know that they may do so. If the shutdown first has to pass through an approval chain, it is worthless in a real incident.
How do you recognize the case? An agent that produces nonsense often does not stand out immediately. Simple thresholds help: unusually many write accesses in a short time, accesses to systems outside the usual pattern, conspicuously long runtimes. Such signals belong in monitoring, not in the memory of individual employees.
What happens afterwards? After the stop, the actual work begins: identify started processes, roll back faulty changes, inform the affected departments. Anyone who has played through this process once quickly notices which logs are missing.
Consequences for the architecture
The fact that a shutdown mechanism is necessary has repercussions for the design of agent systems. An agent that writes every change immediately and permanently is difficult to stop. An agent that creates drafts which a human or a second review step approves can be halted at any time without leaving an inconsistent state behind.
The same applies to fault tolerance. Automatic restarts make sense for classic batch jobs, but with agents they can become a problem – namely when they override a deliberate stop. The restart logic should therefore respect a central lock instead of blindly starting anew.
And finally: agents need to be inventoried. Many companies today do not know exactly how many automated helpers are already running along in development environments, CMS instances or back-office processes. What is not recorded cannot be switched off either.
Assessment
For web and software projects, this shifts a piece of responsibility forward: the question of how an agent is stopped and its state cleaned up belongs in the concept phase, not in operations. Concretely, this means setting up agents with their own identity, narrowly scoped rights and complete logging, and designing processes in such a way that an abort remains tolerable.
Anyone who establishes these foundations can gradually expand automation without having to renegotiate the risk with every new use case.