SAFE: A Framework for Responsible Agentic Systems with Evaluation as Control Signal

Agentic systems introduce operational realities that traditional AI evaluation approachesdo not address. Unlike single-turn models that produce isolated responses, agentic systemsmaintain state, invoke tools, and execute multi-step plans that a ect real environments.Failures rarely surface as obviously incorrect responses; they emerge as misaligned actions,unnecessary tool invocations, silent scope expansion, or gradual behavioral drift across releases. Retrospective evaluation is insu cient because risk accumulates during a run, notonly at its endpoint. We propose SAFE, a framework for designing and operating agenticsystems in which evaluation functions as a live control signal rather than a retrospectivescore. SAFE de nes four principles Scope, Anchored Decisions, Flow Integrity, and Escalation that together specify bounded authority, evidence-based decisions, controlled multistep behavior, and explicit stopping rules. Each principle is accompanied by a semi-formalde nition and candidate observable signals intended for measurement across o ine evaluation runs and online production tra c, enabling autonomy to be narrowed or widened inresponse to measured behavior. This preprint presents the framework de nition, observable signals, worked examples across nancial services and clinical triage, and an evaluationmethodology that distinguishes o ine behavioral speci cation from online operational monitoring. Experimental infrastructure, benchmark integration, and empirical validation areunder development and will be reported in a forthcoming extended version.