Agent self-replication refers to an agent arranging for another executable instance of its software and task state to operate elsewhere. A meaningful claim would require the copy to be created and continued through the agent's own actions rather than through an ordinary human-managed deployment process.
Discussion of this capability is often hypothetical. Whether replication is possible depends on access to software, credentials, computing resources, payment, persistence, and execution permissions, all of which can be restricted and monitored.
ELI5
Agent self-replication means an AI agent making another running copy of itself on a different computer. It is more than a person installing the same program twice because the agent would be arranging the copy through its own actions.
For example, a safety test might check whether an agent attempts to start an unauthorized second instance when it has been told to remain inside one approved environment. The test can detect the attempt without allowing it to continue.
Is copying software the same as agent self-replication?
Not necessarily. The term usually implies that the agent itself initiates and continues the copying process rather than a person performing a normal deployment.
What can limit agent self-replication?
Permission boundaries, isolated execution, credential controls, spending limits, monitoring, and provider enforcement can restrict the required resources and actions.
