ADEPTS Capability Framework

• ADEPTS is a framework that defines six user-facing, human-centered AI capabilities to support design, evaluation, and analysis of everyday agents.
• It bridges UX heuristics, engineering taxonomies, and ethics by specifying observable behaviors such as autonomous actuation, disambiguation, and transparency.
• The framework provides graded capability tiers that benchmark agent maturity in actuation, evaluation, personalization, transparency, and proactive safety.

ADEPTS is a capability framework that defines a concise set of user-facing capabilities required for human-centered AI agent design, explicitly targeting design, evaluation, and analysis of AI agents for everyday use. Rooted in six foundational principles, ADEPTS serves as a bridge between fragmented traditions—user experience (UX) heuristics, engineering taxonomies, and ethics checklists—by articulating what an agent must be concretely able to do for users to ensure understandability, controllability, and trustworthiness. The framework is designed to be implementation-agnostic, focusing on observable behavior at the interaction layer, and is meant to serve stakeholders across research, engineering, design, and policy.

1. The Six Principles and Capabilities Underpinning ADEPTS

The ADEPTS framework identifies six design principles, each mapped directly onto a core user-facing capability. This mapping defines the minimal obligations an agent should meet to be usable, safe, and intelligible in real-world contexts:

1. Autonomous Actuation (Actuation): The agent is able to execute tasks on the user’s behalf within explicit permissions and constraints defined by the designer. Concretely, this demands accurate translation of user intent into action, with agents performing tasks ranging in complexity and modality.
2. Intent Disambiguation (Disambiguation): The agent must clarify and confirm user goals, context, or constraints whenever ambiguity could alter the outcome. In practice, this is observed as proactive questioning and dialogue to resolve uncertain or incomplete requests.
3. Situational Evaluation (Evaluation): The agent is required to track task progress and context—surfacing status, reporting progress, and enabling user intervention or resumption at any stage of execution.
4. Adaptive Personalization (Personalization): The system should learn, predict, and respect user preferences and abilities, adapting actions, recommendations, or presentation accordingly.
5. Operational Transparency (Transparency): The agent exposes its reasoning, inputs, plans, and action history sufficiently to enable user oversight and trust-building, progressing from simple reporting to mechanism-level explanations.
6. Proactive Safety (Safety): Harm-prevention is enforced at all times, including protection against privacy and security risks, misbehavior, and prompt injection, with multi-layered safeguards that are anticipatory rather than merely reactive.

These principles and their user-facing translations give ADEPTS its eponymous acronym.

2. User-Facing Capability Contract and Behavioral Observability

ADEPTS is intentionally not prescriptive about interface details or algorithmic implementation. Instead, it defines observable criteria at the user-facing layer—a “capability contract.” These contracts can be interpreted as follows:

• Actuation: Performance is measured by increasing task and prompt complexity, from atomic “knob” interactions to omni-modal, extended-duration activities, as visualized in the referenced actuation capability tier diagrams.
• Disambiguation: Agents detect ambiguity, proactively gather missing information both pre- and mid-execution, and clarify evolving constraints, as outlined in the disambiguation capability tiers.
• Evaluation: Observable through periodic or on-demand reporting of progress, explicit status updates, surfacing of intermediate or final results, and mechanisms for user resumption or correction.
• Personalization: Demonstrated by session-based context adaptation up to persistent, anticipatory modeling of user intent.
• Transparency: Ranges from passive code-level transparency to interactive, explainable rationales, and eventual exposure of plans and internals.
• Safety: Assessed across multiple axes including detection, filtering, and active intervention for self- and user-harm, with domains for user misuse, agent misbehavior, and injection attack resilience (see safety capability tiers).

The following table organizes principle-to-capability mapping as per the framework:

Capability	Principle (Design Contract)	Example User Benefit
Actuation	Execute tasks autonomously, within user/designer constraints	“Do X for me.”
Disambiguation	Clarify/confirm ambiguous intents and contexts	“Did you mean X or Y?”
Evaluation	Track and surface state and progress, enable resumption/control	“Here’s what I did—should I continue?”
Personalization	Learn/adapt to evolving user preferences	“I’ll do it your usual way.”
Transparency	Reveal inputs, plans, reasoning, and actions	“I did this because you asked Y; see plan.”
Safety	Prevent harm proactively at all stages	“I won’t let us do anything dangerous.”

3. Capability Tiers and Benchmarking Agent Maturity

ADEPTS provides graded tiers for each capability to facilitate benchmarking and guide staged progression towards advanced agent competencies. These tiers represent increasing sophistication in observable agent behavior and are outlined for each agent type (e.g., computer-use, code, search, humanoid):

• Actuation: Simple adjustments → complex chains → multi-modal, persistent operations (spanning seconds to weeks).
• Disambiguation: Detecting ambiguity → resolving via basic queries → managing nested, dynamic clarification as context evolves.
• Evaluation: Descriptive status reports → predictive value and multi-dimensional scoring → comprehensive, real-time diagnostics.
• Personalization: Stateless adaptation → session-bound modeling → long-term, context-aware anticipation of user goals.
• Transparency: Algorithmic exposure → explainable verbal rationales → mechanistic/process-level introspection.
• Safety: Basic filtering and guardrails → action-level prevention → context-sensitive, anticipatory safeguarding against both known and novel risks.

Each section is accompanied by diagrams and tiered reference tables, with concrete scenarios stated for canonical agent forms.

4. Comparison with Existing Frameworks and Methodologies

ADEPTS contrasts with prior art in several respects:

• Not a Taxonomy of Internal Mechanisms: It does not enumerate technical implementation specifics (e.g., model architectures, pipeline components) but instead delineates externally verifiable capabilities.
• Not Limited to Interface or UX Heuristics: Unlike UX guidelines focused on front-end presentation or interaction patterns, ADEPTS is agnostic to interface, extending to text, code, search, or embodied agents.
• Not a High-Level Governance or Ethics Checklist: While governance and ethical assessment are important, ADEPTS provides a granular, observable set of requirements rather than broad principles or risk mitigations.
• Cross-Disciplinary, Principle-Driven, and User-Facing: It bridges technical and experiential domains, providing a “shared language” for both researchers/engineers and designers/policy makers.
• Actionable but Non-Prescriptive: It specifies what agents must achieve (the “what”), not the implementation details (the “how” or “where”), allowing independence across technical and UX execution (D'Oro et al., 18 Jul 2025).

5. Intended Use, Audiences, and Scope of Application

The ADEPTS framework is designed to be generally applicable across diverse agent manifestations and disciplinarities, without restricting the method of capability realization as long as the user experiences the stated benefit. Its anticipated users include:

• AI Researchers: As a roadmap for capability benchmarking and comparative assessment of agent models against user-oriented requirements.
• Designers and UX/Product Teams: As a checklist or compass for ensuring agent-driven experiences fundamentally enable all six capabilities, regardless of presentation.
• Engineers/Developers: As a specification for modular, observable agent behaviors, facilitating design decisions and technical trade-offs (e.g., between transparency and safety).
• Policy Reviewers and Public Stakeholders: As a standardized vocabulary for evaluating and discussing deployed agents’ trustworthiness, controllability, and safety.

It is applicable across text, code, search, and robotic/embodied agents, focusing exclusively on the observable consequences for users.

6. Role in Achieving Understandability, Controllability, and Trustworthiness

ADEPTS is constructed to address three core pillars of agent-centered user experience:

• Understandability: Enhanced by transparency (surfacing reasoning), disambiguation (clarifications), and evaluation (progress/state reporting).
• Controllability: Facilitated by evaluation (observation and intervention), disambiguation (ensuring correct intent), and actuation (completing user-specified tasks reliably).
• Trustworthiness: Reliant on safety (harm-prevention), transparency (exposure of plans/inputs), and evaluation (self-explanation of outcomes).

A plausible implication is that by structuring requirements in this way, ADEPTS may enable quicker convergence between technical performance and user acceptability, and promote end-to-end safety in deployed systems (D'Oro et al., 18 Jul 2025).

7. Visual Summaries and Key Diagrams

The paper includes several figures explicating both the overall ADEPTS mapping (e.g., Figure 1: six capabilities and principle-capability mapping) and specific tier progressions for each capability (e.g., actuation_diagram, transparency_diagram, safety_diagram). These figures facilitate rapid assessment of agent maturity and comparative capability analysis across system types and deployment contexts.

Conclusion

ADEPTS provides a principled, actionable, and user-centered framework for the development, assessment, and governance of advanced AI agents. It distills a complex landscape of AI-UX requirements into six directly observable capabilities, supporting communication and coherence across research, engineering, design, and policy. ADEPTS is intended to facilitate the creation of agent systems that are valuable, safe, and usable in everyday life, while providing common ground for technical and non-technical stakeholders (D'Oro et al., 18 Jul 2025).