Digital twins are migrating from mirrors to makers: as AI gives them agency, tighter coupling and evolving models can turn decision‑support tools into systems that reshape the realities they model, creating both new capabilities and risks of performative lock‑in.

Main Finding

Burr et al. (Jan 2026) develop a three‑dimensional taxonomy for “agentic” digital twins (DTs) that integrates AI-driven autonomy with systems theory and performative prediction. By decomposing agentic DTs along (1) locus of agency (external, internal, distributed), (2) tightness of coupling to the physical target (loose, tight, constitutive), and (3) model evolution (static, adaptive, reconstructive), they generate a 27‑configuration space and highlight nine illustrative configurations grouped into three clusters (“The Present”, “The Threshold”, “The Frontier”). The paper’s central conceptual claim is that agentic DTs can exercise performative (ontological) power: beyond representing systems, they can co‑constitute what is measured and governed, producing feedbacks that create path dependence, lock‑in, or even redefinition of ontologies.

Key Points

• Taxonomy: three binary/ternary dimensions produce 27 formal configurations, not all equally plausible or desirable:

  • Agency: External (E), Internal (I), Distributed (D)
  • Coupling: Loose (L), Tight (T), Constitutive (C)
  • Evolution: Static (S), Adaptive (A), Reconstructive (R)
  • Compact notation: ⟨Agency, Coupling, Evolution⟩ (e.g., ⟨I, C, A⟩).

• Nine illustrative configurations (representative, not exhaustive):

• Computational Tool ⟨E, L, S⟩ — conventional DT as instrument.
• Adaptive Monitor ⟨E, T, A⟩ — real‑time, learning monitor used by external agents.
• Active Steering ⟨I, T, A⟩ — embedded agent uses DT to steer the target.
• Symbiotic ⟨D, T, A⟩ — behaviour emerges from tight human–DT coupling.
• Governor ⟨I, C, A⟩ — constitutively coupled, adaptive, internally agentic; strong performative lock‑in risk.
• Swarm ⟨D, T, S⟩ — distributed coordination via static DT representations.
• Worldbuilder ⟨E, L, R⟩ — external actor reconstructs ontologies in a sandboxed manner.
• Voyager ⟨I, T, R⟩ — internally agentic DT that autonomously redefines categories.

• Reconstructive Assemblage ⟨D, C, R⟩ — maximal distributed, constitutive, reconstructive system.

• Three clusters of development:

  • The Present (configurations 1–3): existing or emerging DTs (tools, monitors, steering).
  • The Threshold (4–6): emergent properties, distributed agency, constitutive coupling (high governance relevance).
  • The Frontier (7–9): reconstructive capability, ontological reconstitution (speculative but technically conceivable).

• Performativity and ontological power:

  • DTs can shape the realities they model (measurement affects the target).
  • “Governor” exemplifies systems that define what they measure, potentially creating self‑validating realities and foreclosing alternatives.
  • More advanced reconstructive systems can change the underlying ontology (how entities and relations are defined), with possible epistemic inscrutability.

• Dynamics and transitions:

  • The paper provides notation for progression paths and for capturing drivers and barriers of transitions (technological pressures, design choices, governance interventions).
  • Traffic navigation examples illustrate how the same domain manifests different performative dynamics across configurations.

• Normative and methodological framing:

  • Authors adopt a pragmatic constructivist stance: physical systems exist independently, but constituting them as particular technical objects (with specific measurements and interventions) is sociotechnical and performative.
  • The taxonomy itself is acknowledged to be performative: categories influence design and governance.

Data & Methods

• Methodological approach: conceptual/theoretical taxonomy built from synthesising:
  • Systems engineering concepts of coupling and interdependence,
  • AI/agent frameworks (multi‑agent, embedded autonomy, LLM orchestration),
  • Philosophical/pragmatic constructivism and performative prediction literature.
• Formal output: 3 × 3 × 3 coordinate space (27 configurations) and a compact notation for configurations and paths.
• Illustrative grounding: a series of case examples (not large‑scale empirical datasets), notably multiple treatments of traffic navigation systems showing how performativity unfolds across configurations.
• Analytical tools: thought experiments, configuration clustering, progression/path notation, and engagement with existing formal tools of performative prediction theory to reason about feedbacks and lock‑in.
• Limitations acknowledged by authors: primarily a conceptual paper (not empirical validation), taxonomy choices may themselves shape development, and some configurations are speculative or technologically challenging.

Implications for AI Economics

• Market structure and incumbency

  • Performativity can create strong first‑mover advantages: DTs that shape measurement/metrics can bias markets toward providers that control those metrics (measurement rents).
  • Constitutive coupling and distributed assemblages may entrench incumbents or platform owners through data‑dependent lock‑in and emergent standards.

• Valuation, investment risk, and uncertainty

  • Reconstructive DTs introduce novel epistemic uncertainty: product/service value becomes entangled with the DT’s ability to redefine market categories or create demand by reshaping perceptions.
  • Investors face higher model risk and tail risks where DTs materially alter the fungibility or comparability of assets and services (e.g., insurance pricing, asset valuation).

• Product design, competition, and market power

  • Strategic incentive to design DTs that shape demand (performative pricing): firms may deliberately tune DTs to favour their offerings, raising regulatory concerns.
  • Interoperability and openness become central pro‑competitive levers to prevent proprietary ontological lock‑in.

• Externalities and social welfare

  • Feedback loops (e.g., routing/navigation DTs) can generate negative externalities (congestion spillovers, spatial displacement, distributive harms) that are hard to internalise if DTs are privately governed.
  • Constitutive systems can narrow policy or consumer choices by redefining feasible actions; welfare calculations must account for loss of alternatives and epistemic harms.

• Labour, contracting, and skill demand

  • Automation of decision‑making via internal/distributed DTs shifts demand toward maintenance, governance, auditing, and higher‑order design skills (and away from fine‑grained operational roles).
  • New contractual forms and liability regimes will be required where DTs co‑constitute outcomes (who is accountable for ontological choices?).

• Regulation, governance, and public infrastructure

  • Near‑term priority: monitor and regulate configurations in “The Threshold” (e.g., ⟨I, C, A⟩ Governor) because they are technologically reachable and pose significant performative/lock‑in risks.
  • Policy tools: mandated disclosure of model evolution and coupling characteristics, auditability requirements, human‑in‑the‑loop constraints, standards for interoperability, and sunset/rollback mechanisms to avoid irreversible ontological shifts.
  • Public provision: for domains with major externalities (transport, energy, health), publicly governed DT infrastructure or neutral intermediaries could limit biased performativity and enable social welfare–oriented design.

• Measurement, metrics, and economic modelling

  • Traditional economic models assume exogenous measurement. Agentic DTs violate that assumption: endogenous measurement processes must be modelled (performative prediction).
  • Empirical work should incorporate counterfactual and causal inference methods robust to feedback (e.g., experiments, randomized allocation of DT interventions, instrumental variables, market simulations).

• Policy research and empirical agenda

  • Develop empirical diagnostics for performativity (detecting when a DT is beginning to constitute its target).
  • Stress‑test markets under scenarios of constitutive coupling and reconstructive DTs (agent‑based simulation, market design experiments).
  • Design regulatory experiments (sandboxes) for reconstructive capabilities with strong oversight and rollback.

Practical recommendations (concise) - Prioritise governance and auditability for systems that exhibit internal agency + constitutive coupling or adaptive + reconstructive evolution. - Promote interoperability and data portability to reduce measurement‑driven lock‑in. - Require transparency about model adaptability and ontological changes; mandate public reporting for DTs used in essential infrastructure. - Fund empirical research and simulation exercises to quantify market and distributional risks from performative DTs.

Overall, the taxonomy gives AI economists a structured vocabulary and conceptual tools to analyse how autonomous, data‑rich DTs could alter incentives, market structure, valuation, and welfare by making measurement and modelling active components of economic systems rather than passive observers.