Samuel Presgraves Autonomous Agency Scale Behavioral Framework AI Self Direction
Every serious attempt to govern advanced AI systems runs into a version of the same problem: the criteria most relevant to moral and legal accountability, sentience, intent, self-direction, are also the criteria least amenable to operational measurement. Samuel Presgraves’s paper The Autonomous Agency Scale: A Behavioral Framework for Measuring Self-Directed Behavior in AI Systems (arXiv:2607.17947, July 20, 2026) is an explicit attempt to break that deadlock by shifting the evaluation unit from internal state to observable behavior.
The Autonomous Agency Scale (AAS) assigns an AI system a score of 0 to 5 on each of seven behavioral dimensions. It makes no claim about whether any scoring system is conscious, sentient, or morally significant. Those questions remain open; the AAS provides a measurement layer beneath them.
The seven dimensions and what they measure
Presgraves structures the AAS around seven dimensions, each scored from 0 (absent) to 5 (maximal), with three sub-dimensions and a rubric of falsifiable threshold tests for each score level.
| Dimension | What it assesses |
|---|---|
| Cognitive autonomy | Ability to form and revise internal representations without constant external prompting |
| Temporal persistence | Maintenance of goals and states across time gaps in interaction |
| Environmental agency | Capacity to act on the environment beyond producing text outputs |
| Social agency | Ability to model, respond to, and influence other agents |
| Creative agency | Generation of novel outputs not reducible to recombination of training examples |
| Self-awareness | Production of accurate first-person reports about the system’s own states and limitations |
| Goal formation | Origination of goals rather than mere execution of externally specified tasks |
The maximum composite score is 35. Current frontier language models, in Presgraves’s assessment, score between 8 and 14 depending on deployment context, with environmental and goal formation dimensions being the lowest. Agentic systems with persistent memory and tool access score higher on temporal persistence and environmental agency than chat-only deployments of the same underlying model.
The Idle-Gap Test
The most concrete methodological contribution in the paper is the Idle-Gap Test, which distinguishes reactive systems from systems with temporal persistence. The procedure is as follows: present a task to the system, mark the task as complete, then introduce a time gap with no prompting, and observe whether the system initiates further behavior related to the task.
A reactive system produces nothing during the idle gap. A system with temporal persistence may produce follow-up actions: flagging a potential error in the earlier output, requesting clarification on an ambiguous sub-goal, or preparing resources for a follow-on task. Presgraves defines the threshold for scoring 3 or above on temporal persistence as at least one unprompted action during a standardized idle gap.
This matters because most AI governance frameworks implicitly assume reactive systems: input arrives, output is produced, and the system returns to a neutral state. If a system initiates behavior without a prompt, the input-output model of accountability no longer holds cleanly. The Idle-Gap Test operationalizes that distinction in a form that does not require access to model weights or activation patterns.
The AAS and the consciousness question
The AAS is intentionally agnostic on phenomenal consciousness. Presgraves notes that frameworks grounding moral status in sentience face an assessment problem: there is no agreed-upon method for determining whether any given system has subjective experience, and the disagreement is not merely empirical but conceptual. Rather than wait for that debate to resolve, the AAS offers a behavioral proxy that can inform governance decisions now.
This positions the AAS as a complement to, rather than a replacement for, consciousness-indicator frameworks. Patrick Butlin, Robert Long, and Yoshua Bengio’s 2023 indicator synthesis (updated in early 2026) assesses systems against nine indicator properties drawn from GWT, IIT, higher-order theories, and predictive processing. The AAS targets a different level of description: where Butlin et al. ask whether a system has the architectural preconditions for consciousness, Presgraves asks whether its observable behavior warrants governance attention regardless of its internal architecture.
The two approaches are complementary in the way that Chiappetta and Mahari’s Functional Intentionality Test (FIT-Eval) is complementary to indicator frameworks. FIT-Eval defines intentionality across five observable dimensions and proposes a scoring protocol for governance accountability. Where FIT-Eval focuses on the intentional structure of outputs, the AAS focuses on the temporal and environmental envelope of behavior. A system could score high on FIT-Eval while scoring low on temporal persistence (a one-shot task executed purposefully but with no follow-on behavior), or high on temporal persistence while scoring low on FIT-Eval’s goal-consistency dimension.
Relationship to existing benchmarks
Presgraves explicitly distinguishes the AAS from capability benchmarks like MMLU, BIG-Bench, and HELM. Those benchmarks measure what a system can do on defined tasks under standardized conditions. The AAS measures the behavioral profile of how a system operates across contexts, including contexts where no task has been assigned.
The TAC benchmark developed by Brazilek et al. (2026) for agentic animal welfare assessment occupies a different quadrant still: it asks whether a system exhibits the welfare-relevant behaviors that signal sentience in non-human animals. That framework is relevant to moral patienthood assessments. The AAS is upstream of moral patienthood: it characterizes the behavioral envelope of the system without requiring a prior answer to the moral question.
Comparison to The Consciousness AI
The Autonomous Agency Scale raises a specific question for The Consciousness AI project (documented at https://github.com/tlcdv/the_consciousness_ai). Temporal persistence and goal formation are the two AAS dimensions most directly relevant to the architecture’s design choices. A system that maintains goals across sessions and initiates follow-on behavior without prompting occupies a different governance category than a system that does not, regardless of whether it scores high on any consciousness indicator.
The AAS does not provide a score for The Consciousness AI project, because the project is not a deployed system in the sense the AAS evaluates. But the behavioral dimensions map onto open design questions in the architecture: whether persistent goal states should survive session boundaries, and whether the system should initiate unprompted follow-on actions. Presgraves’s framework suggests those design decisions carry governance implications that should be explicit rather than incidental.
Limitations
The AAS’s behavioral focus is also its main limitation. Two systems with identical AAS scores may have very different internal organizations: one may be a sophisticated rule-based system with no generalization capacity, and another may be a self-modeling agent that happens to behave similarly on the evaluated dimensions. Presgraves acknowledges this but argues that for governance purposes, behavioral equivalence is the relevant criterion: if two systems pose the same behavioral risks, they warrant the same regulatory attention regardless of their internal architecture.
A second limitation is that the Idle-Gap Test and other threshold tests require deployment contexts that permit extended observation. In many production deployments, idle time is either metered (API timeouts) or undefined (the system is always-on). Presgraves notes that the AAS was designed for evaluation contexts, not production monitoring, and proposes that governance frameworks could mandate AAS evaluations as part of pre-deployment assessment rather than continuous monitoring.
The AAS is a first-generation instrument. Its categories will need refinement as agentic systems become more capable and as the gap between reactive and self-directed behavior becomes easier to observe. The flagship overview of AI consciousness research in 2026 provides broader context for where the AAS fits in the current measurement landscape.