The survey's central organising choice is a division between two philosophical traditions in building agents that act in the world. Symbolic AI agents plan and reason using explicit representations: logic rules, knowledge graphs, formal ontologies. They are interpretable by construction, because every decision traces back to a named rule or fact. Their weakness is brittleness: they break when reality deviates from the schema their designers encoded.
Neural AI agents, including LLM-based systems, learn behaviour from data and generalise across situations no rule could anticipate. Their strength is flexibility; their weakness is opacity. When a neural agent takes an action, the reasoning is distributed across billions of parameters with no clean audit trail. For high-stakes domains, this is not an engineering inconvenience but a regulatory and ethical barrier.
The survey's PRISMA-guided methodology covers both paradigms across six domains, enabling cross-cutting comparisons that single-paradigm reviews miss. The authors are careful not to declare a winner. Instead, they document where each approach has established genuine advantages and where the gaps remain large.
Flexible, context-aware, opaque
Learn behaviour from data; generalise across novel situations; handle natural language and ambiguity naturally. Weak on auditability, formal guarantees, and predictable failure modes. Current frontier: LLM-based agents with tool use and memory.
Interpretable, brittle, verifiable
Operate on explicit rules and knowledge representations; every decision is traceable; support formal verification. Weak on generalisation and robustness to out-of-distribution inputs. Current frontier: knowledge-graph-augmented planning systems.
The most striking finding across the domain analysis is how differently the paradigm trade-offs land depending on the deployment context. In robotics, symbolic approaches remain dominant for structured manipulation tasks where formal safety constraints are non-negotiable, while neural approaches have taken over in unstructured environments such as household assistance, where rigid rule sets cannot anticipate every configuration a robot might encounter.
Healthcare is the domain where the stakes of explainability are highest, and the survey documents the resulting conservatism. Clinical decision support systems lean heavily symbolic even when neural alternatives would outperform them on held-out benchmarks, because regulatory frameworks and clinician trust require auditable reasoning chains. The authors note that the field is experimenting with hybrid architectures where a neural module generates candidate recommendations and a symbolic module filters them against clinical guidelines, but no approach has yet reached broad deployment.
Finance shows the opposite dynamic: neural agents have penetrated trading and risk assessment more rapidly than almost any other sector, driven by competitive pressure to exploit market signals that resist formalisation. The regulatory response is still catching up, and the survey documents a gap between deployed capability and available oversight mechanisms that the authors flag as a systemic risk.
Robotics
Hybrid approaches dominate; symbolic for safety constraints, neural for unstructured perception and manipulation.
Healthcare
Explainability requirements favour symbolic; hybrid recommendation filtering is the emerging pattern.
Finance
Neural agents deployed fastest; regulatory oversight lags capability; systemic risk from opaque decision chains.
Education
Personalised tutoring favours neural; curriculum sequencing benefits from symbolic structure and curriculum standards.
Smart Infrastructure
Real-time control requires formal guarantees; symbolic planning with neural perception preprocessing is the standard pattern.
Social Systems
Multi-agent coordination, policy simulation; both paradigms active; accountability and fairness remain open problems.
Across all domains, the survey identifies two recurring themes in what remains unsolved. The first is trustworthiness at deployment scale. Most agentic AI research measures performance on benchmarks designed by the researchers themselves, against tasks with known ground truth. The survey documents a consistent drop in reported performance when systems are evaluated on independent benchmarks or in real operational environments, a gap the authors attribute to overfitting to evaluation conditions rather than generalisation to deployment requirements.
The second is the explainability-capability trade-off at the architecture level. Hybrid systems that combine symbolic and neural components are the most promising direction for both, but the field lacks standard interfaces between the paradigms. A neural module's output is a probability distribution or token sequence; a symbolic module expects logical predicates. Bridging this representation gap without degrading either component's performance is an open engineering challenge that the survey documents but does not resolve.
The governance dimension receives its own section, and the authors are direct: most current agentic AI deployments operate without meaningful oversight mechanisms. This is not primarily a technical failure. The regulatory frameworks, professional standards, and audit protocols that would constitute meaningful governance do not yet exist for most domains. The survey's implicit argument is that producing them should be treated with the same urgency as producing better models.
The field's current trajectory, improving neural agent capability while deferring explainability and governance, is not sustainable for the domains where agentic AI has the most transformative potential. Healthcare, infrastructure, and policy applications require trustworthy reasoning chains, not just accurate outcomes. The authors' case for hybrid symbolic-neural architectures is, at bottom, a case for treating interpretability as a first-class design requirement rather than a post-hoc analysis task.