The Recoverable Agent: Immune-like defence, sleep-like maintenance and bounded autonomy

AI agents can execute consequential actions faster than organisations can investigate and approve a response. This paper proposes a conceptual architecture in which compartmentalised agents operate alongside two complementary processes: continuous, immune-like protection and periodic or event-triggered, sleep-like maintenance. The protective process enforces delegated authority, detects selected deviations and isolates affected compartments. The maintenance process suspends external action, examines memory provenance, removes disposable state, restores trustworthy records, rebuilds suspect components and tests readiness before reactivation. Both are supervised by a protected control plane outside the working agent's authority. Essential work may transfer to clean reserve compartments, but potentially compromised agent state does not migrate by default. The proposal connects technical controls with GCTI's proposed Red-Team Assessor (RTA) and AI Security Operator (AISO) roles. Its central claim is conditional: pre-authorised containment can remove a human approval step from the immediate response path, but cannot eliminate detection uncertainty, guarantee instantaneous repair or reverse completed harm. Biological sleep and immunity provide a functional analogy, not evidence that software sleeps, experiences fatigue or becomes conscious. The paper distinguishes autonomous immunity from autoimmunity, treats harmful defensive reactions as a first-class risk, and sets out falsifiable hypotheses and a prospective evaluation protocol. No implementation, experiment or measured effectiveness is reported. The contribution is an integrated, testable account of bounded action, recoverable state and professional accountability, rather than a claim to have invented self-healing computing or achieved invulnerability. Final conceptual research report. Not peer reviewed. No implementation, experiment or measured effectiveness is reported. Declared interests and AI-assistance disclosures appear in the paper.