AI agents at the brain-computer interface: separating inference from control

In medicine, AI agents are moving from generating text to executing actions, making uncertainty from upstream decoders a control problem. We studied this at the brain-computer interface using 1,065 episodes from 47 people with amyotrophic lateral sclerosis and five language models. Prompting agents with reconstructed decoder confidence never reduced unfaithful execution below a deterministic gate at matched coverage; two models were significantly worse. Apparent safety gains of up to 22 percentage points reflected acting less often, sometimes through invalid tool calls rather than explicit abstention. A post-hoc fair-information test gave ten models the same command vocabulary as a deterministic resolver. No direct agent arm improved on the resolver's risk-coverage frontier, but a hybrid architecture in which models proposed semantic corrections and an external gate retained admission authority extended coverage beyond the resolver in five of ten models without observed unfaithful executions. These results separate inference from control in agentic neurotechnology. Funding: A.G. and E.K. were supported in part by the Clinical and Translational Science Awards (CTSA) grant UL1TR002541 from the National Center for Advancing Translational Sciences, through the Harvard Catalyst | The Harvard Clinical and Translational Science Center Pilot Award Program. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.