Circuit breakers (AI agents)
A programmed safety mechanism that automatically stops an AI agent from running when it detects predefined anomalies, such as infinite loops, goal deviations, mass access to data or attempts at mass exchanges of information. The AEPD guidelines on agentic AI address it together with hard step limits (section VII.F, p. 72).
It is the agent’s automatic brake: it acts when what the agent does goes outside what was planned, without waiting for someone to notice.
Which obligations it carries
It is a best practice, not a legal obligation: it is one of the measures for controlling the agent by design proposed by the guidelines. It rests on obligations that are legal: processing personal data in a manner that ensures appropriate security, including protection against unauthorised or unlawful processing and against accidental loss, destruction or damage (Article 5(1)(f) of Regulation (EU) 2016/679), and implementing appropriate technical and organisational measures to ensure a level of security appropriate to the risk (Article 32(1)), both applicable from 25 May 2018.
What it is not
It is not human oversight: the circuit breaker acts without anyone intervening, and human oversight is a separate measure that the guidelines address on its own. Nor is it the same as the step limit it is grouped with: the limit stops a task when it reaches a maximum number of steps even if nothing looks anomalous, and the circuit breaker stops it when something does. And it is not a promise that the agent will not fail: the guidelines start from accepting the possibility of failure as an element of governance (section VII.A, p. 55).
The nuance almost nobody captures
A circuit breaker only detects what was defined in advance as an anomaly. That is why the guidelines describe it by type of anomaly — loops, goal deviation, mass access, mass exchange — rather than as a generic switch, and why the list of anomalies is part of the agent’s design, not a later adjustment.