Framed Autonomy in Autonomous Systems

• Framed autonomy is a conceptual framework that bounds and quantifies agent capabilities relative to a defined reference frame.
• It employs layered architectures and multi-metric quantification (e.g., reliability, responsiveness) to assess and certify autonomy levels.
• Hybrid governance models integrate technical, epistemic, and social controls to limit decision power and ensure safe autonomous operations.

Framed autonomy denotes a class of formalisms, architectures, and governance models in which the autonomy of artificial or human agents is explicitly bounded, instantiated, and measured relative to a particular reference structure or “frame.” Across domains from cyber-physical systems and critical infrastructure to decision-theoretic models of agency and multi-agent AI, framing autonomy serves to clarify, quantify, and control the degree and scope of self-governance, decision authority, and adaptive behavior in complex, distributed, or high-stakes environments. Core dimensions include specifying structural and operational boundaries for autonomy, decomposing autonomy into layers or levels, introducing quantitative benchmarks or certification protocols, and surfacing the formal and social power allocations that determine who (or what) acts, chooses, or retains the right to decide.

1. Formalization and Layered Architectures of Framed Autonomy

Framed autonomy is foundationally defined by structuring autonomous capabilities and permissions with respect to explicit boundaries, architectural levels, or agent-environment decompositions.

Embedded, Layered Autonomy in Engineered Systems:

In the context of spacecraft and complex cyber-physical platforms, autonomy-at-levels formalizes the design principle that autonomy functionality should not be centralized in a single controller. Instead, outer “autonomy loops” wrap around inner control loops at every point of a systems engineering hierarchy—components, assemblies, subsystems, and full systems (Baker et al., 19 Aug 2025). Each autonomy loop adapts the setpoints, modes, or behaviors of its corresponding control loop based on sensed local environmental conditions, goals, and crosstalk messages. This nested architecture is described by equations for inner and outer closed-loop transfer functions, e.g.,

$$G_c(s)=\frac{C(s)P(s)}{1+C(s)P(s)},\quad G_a(s)=\frac{A(s)}{1+A(s)G_c(s)},$$

and can be composed into the overall system behavior.

Meta-Modeling in Multi-Agent IoT/CPS:

The Autonomy Model and Notation (AMN) introduces UML-style meta-models for IoT/CPS agents where agents are composed hierarchically; interfaces (sensors, actuators) have typed, constrained connectivity; and agents’ autonomy is parameterized by a suite of explicitly modeled properties (social self-concept, ethical stance, rules, goals, states) (Janiesch et al., 2020). Constitutive characteristics and meta-model constraints formalize how autonomy is instantiated, measured, and managed.

2. Quantitative Metrics, Levels, and Degrees of Autonomy

Quantitative assessment in framed autonomy differentiates between structure (level of autonomy) and operational performance (degree of autonomy).

Multi-metric Quantification:

The framework of Gyagenda and Roth (Gyagenda et al., 2023) maps classical human job characteristics (skills, responsibility, effort) to robot-task metrics:

• Requisite Capability Set: Mandatory functions the system must possess.
• Reliability: How closely empirical error variances per capability meet reference bounds, $$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$.
• Responsiveness: Speed of reaction relative to reference time, $C_{res,i} = \frac{t_{ref,i}}{t_{act,i}}$.

Levels and Degree of Autonomy:

• Level of Autonomy (LoA): Ordinal category indicating which capabilities meet (or miss) reliability and responsiveness thresholds (from external control, through partially autonomous, to full autonomy: LoA 0–4).
• Degree of Autonomy (DoA): Continuous ratio quantifying “by how much” thresholds are exceeded, e.g.,

$$DoA = n^2 \left[\sum_{i=1}^n \frac{1}{C_{rel,i}C_{res,i}}\right]^{-1},$$

enabling side-by-side ranking of systems at the same LoA.

This multi-factorial view is complemented by adaptively assigning autonomy certificates, using task-based pass/fail levels (operator, collaborator, consultant, approver, observer), and deploying assisted-evaluation metrics as in the autonomy level definition of (Feng et al., 14 Jun 2025).

3. Reference Frames, Framing, and the Science of Agency

A central theoretical result in autonomy research is that agency and autonomy are not absolute; they are defined relative to a chosen reference frame.

Frame-dependence of Agency:

Agency (and, by extension, autonomy) must be measured with respect to a clearly specified “frame”—boundaries between system and environment, choice of causal variables, goal specification, and reward structures (Abel et al., 6 Feb 2025). Even normativity (goal-directedness) and adaptivity are frame-sensitive; altering the state abstraction, action set, or reward mapping shifts measured agency:

• Agency in RL: $$Agency(F) = (V^*(s_0) - V_{\pi_0}(s_0))/(max_\pi V^*(s_0) - min_\pi V^*(s_0))$$, where $F$ is the MDP frame.

All claims about “autonomy” must specify—sometimes axiomatize—the frame under which autonomy is operationalized, a principle also echoed in practical agent evaluation and interaction design.

4. Oversight, Bounded Autonomy, and Hybrid Governance

Framed autonomy systematically limits and allocates decision authority using context-sensitive constraints and state-dependent oversight modes.

Governance Automata and Contextual B(S):

Bounded autonomy is operationalized as a mapping $B: X \to 2^M$ (from world state to sets of allowed machine actions) (Sharma et al., 16 Mar 2026). Four archetypal oversight modes partition the operational state-space:

• Fully Automated (FA): Autonomous action under routine, low-risk, low-consequence, low-complexity conditions.
• Human-on-the-Loop (HOTL): Human can interrupt or override proposed agent actions.
• Human-in-the-Loop (HITL): Agent actions require explicit human approval.
• Human-in-Command (HIC): Humans set goals, constraints, may review AI-generated policies at checkpoints.

The partition is formally driven by risk, complexity, and consequence assessments, and oversight transitions can be implemented as event-triggered automata. Compliance with EU AI Act, ISO/IEC safety/controllability standards, and sector-specific crisis management protocols relies on demonstrably respecting these bounded regions.

5. Bounding Decision Power and Mechanical Selection Governance

Beyond action-level filtering and alignment, framed autonomy decomposes and bounds agentic power itself.

Autonomy as a Vector and Mechanical Primitives:

In regulated settings, autonomy is decomposed into a sovereignty vector $\mathbf{A} = (A_{cog},A_{sel},A_{act})$ (Rodriguez et al., 16 Feb 2026):

• Cognitive Autonomy ($A_{cog}$): Internal reasoning capacity.
• Selection Autonomy ($A_{sel}$): Power to generate/frame candidate choices.
• Action Autonomy ($$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$0): Power to execute choices externally.

High-stakes systems retain maximal $$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$1, but mechanically bound $$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$2 and $$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$3 through primitives:

• External Candidate Expansion and Freezing Layer (CEFL): Displaces candidate generation.
• Governed Reducer: Applies policy filters, variance clamping, and buckets for diversity and compliance.
• Commit-reveal entropy isolation: Prevents entropy probing and deterministic outcome capture.
• Rationale validation and fail-loud circuit breakers: Enforces balanced presentation and audible error-signaling.

Operational metrics (selection concentration, narrative diversity, governance activation cost, failure visibility) quantify residual selection influence and governance effectiveness.

6. Social, Epistemic, and Relational Framings of Autonomy

Framed autonomy encompasses not only technical, but also epistemic and social considerations.

Epistemic Framing and Paternalism:

Experiments in decision-maker autonomy demonstrate that autonomy is frequently granted or withheld based on the perceived knowledge or competence of the agent, with knowledge functioning as the currency for autonomy rights (Grossmann, 2024). Policymakers frame interventions not only through rules but by (non-)provision of information, with implications for soft/hard paternalism and the risk of strategic omissions.

Relational Sovereignty:

Assistive technology research critiques the reduction of autonomy to independence and proposes a shift to “relational sovereignty,” which centers the user’s recognized authority to select between independence and interdependence—explicitly reframing autonomy from “can they do it alone?” to “do they get to decide the mode and terms of assistance?” (Jang et al., 8 Mar 2026). The Relational Sovereignty Matrix $$C_{rel,i} = \frac{\sigma^2_{ref,i}}{\sigma^2_{act,i}}$$4 formalizes this set of options, and sovereignty-centered design interventions surface power asymmetries embedded in technology and policy.

7. Practical Guidelines, Limitations, and Cross-Domain Synthesis

Framed autonomy frameworks typically advocate scoping models to domain-specific constructs, embedding autonomy “templates” within UML/BPMN-style notations for agent-based systems (Janiesch et al., 2020), and selecting evaluation metrics with explicit reference to context and involved parties (Feng et al., 14 Jun 2025). Current frameworks have limitations:

• Security and ethics are often modeled merely as aggregate flags rather than through fine-grained policies or engines.
• Conceptual breadth of meta-models can limit domain focus; tailoring is needed.
• Integrative verification/validation for emergent behaviors in nested, multi-level systems remains a significant challenge (Baker et al., 19 Aug 2025).

Nevertheless, across platforms (IoT, AI, social systems, critical infrastructure), framed autonomy offers principled structures for deliberative calibration, assessment, and governance of autonomous behavior in distributed, interactive, and socially loaded environments.