AgREE: Agentic Reasoning for Emerging Entities

• AgREE is an agent-based framework that dynamically incorporates emerging entities into knowledge graphs through iterative retrieval and triple construction.
• The framework leverages causal reasoning, active inference, and system-centric governance to adapt and validate novel information in open, dynamic settings.
• Empirical benchmarks show enhanced performance in open-domain KGC and enterprise reasoning with improved accuracy, efficient multi-hop reasoning, and robust governance.

Agentic Reasoning for Emerging Entities (AgREE) denotes a line of work on agentic reasoning in open-world, dynamic environments where salient entities, relations, and constraints are not fixed in advance. In its narrowest and most explicit usage, AgREE is an agent-based framework for open-domain knowledge graph completion (KGC) on emerging entities that are absent from the knowledge graph and typically unseen during LLM pretraining (Zhao et al., 6 Aug 2025). In a broader research usage, the term is applied to system-level agentic wholes, dynamic 6G agents, uncertain streaming events, enterprise entities, and institutionally governed agent collectives whose capabilities arise through interaction, retrieval, memory, communication, and governance rather than from a single static model (Miehling et al., 28 Feb 2025).

1. Scope and meanings of “emerging entities”

The literature does not treat “emerging entities” as a single ontology. In open-domain KGC, they are entities not present in the KG, or appearing after a model’s pretraining cutoff, especially new and unpopular entities in daily news (Zhao et al., 6 Aug 2025). In the systems-theoretic account of agentic AI, they are system-level agentic wholes whose functional agency arises from interactions among simpler agents, tools, humans, and environments (Miehling et al., 28 Feb 2025). In 6G reasoning-native communication, the relevant entities are dynamic, heterogeneous, and resource-bounded agents whose internal policies and world models evolve and are non-identical (Seo et al., 19 Feb 2026). In multimodal reasoning benchmarks such as GeoBrowse, the analogous objects are salient but under-represented or intentionally hidden actors, institutions, places, and events that must be surfaced from weak signals and external verification (Geng et al., 5 Apr 2026). In enterprise reasoning, they include new suppliers, novel contract types, and newly introduced workflow states (Wang et al., 14 May 2026).

Context	Emerging entity	AgREE emphasis
Open-domain KGC	Entities absent from the KG and typically unseen during LLM pretraining	Iterative retrieval and triple construction
Systems-theoretic agentic AI	System-level agentic wholes	Emergent cognition, causal reasoning, metacognition
6G reasoning-native communication	Dynamic, heterogeneous, resource-bounded agents	Belief alignment and coordination drift prevention
GeoBrowse	Obfuscated or long-tail places, institutions, and events	Multi-hop verification from weak visual cues
Enterprise reasoning	New suppliers, contract types, workflow states	Ontology-grounded, auditable multi-hop reasoning

This plurality is central rather than accidental. AgREE is consistently concerned with settings in which novelty is operationally significant: new entities appear after training, entity names may be absent or obfuscated, interaction topologies change over time, and system behavior cannot be inferred from isolated model capability. A plausible implication is that AgREE is best understood as a family of methods for discovering, representing, validating, and governing novelty under partial observability rather than as a single benchmark-specific technique (Zhao et al., 6 Aug 2025).

2. Theoretical foundations of agency, emergence, and reasoning

A common foundation is the move from model-centric evaluation to system-centric analysis. The systems-theoretic formulation defines functional agency through three conditions: action generation, outcome model, and adaptation. It also places agency on three hierarchies: action generation from reactive to stateful to epistemic, outcome modeling from association to intervention to counterfactual, and adaptation from contextual to parametric to reflective (Miehling et al., 28 Feb 2025). This allows high functional agency to arise at the level of a system even when individual components are simpler.

Several formal vocabularies recur. Agent–environment interaction is cast as a POMDP with tuple $\langle S, A, T, R, \Omega, O, \gamma \rangle$, with policy $\pi(a_t \mid h_t)$ over histories or epistemic beliefs (Miehling et al., 28 Feb 2025). Causal reasoning is organized by Pearl’s hierarchy and SCMs of the form $M = \langle U, V, F, P(U) \rangle$, with intervention semantics via $p(y \mid do(x))$ (Miehling et al., 28 Feb 2025). Predictive processing and active inference are introduced through variational free energy

$F(q) = \mathbb{E}_{q(z)}[\ln q(z) - \ln p(x, z)],$

and action selection by expected free energy $\mathcal{G}_t$ (Miehling et al., 28 Feb 2025). In the survey literature, agentic reasoning is further stratified into foundational single-agent reasoning, self-evolving agent reasoning, and collective multi-agent reasoning, with a second distinction between in-context reasoning and post-training reasoning (Wei et al., 18 Jan 2026).

Another major foundation comes from AAMAS. “Agentifying Agentic AI” formalizes beliefs, desires, and intentions through modal operators $B_i \phi$, $D_i \phi$, and $I_i \phi$, and treats commitments, communication protocols, mechanism design, and institutional modeling as the scaffolding that turns behavioral autonomy into accountable agency (Dignum et al., 21 Nov 2025). Deontic operators $O\phi$, $\pi(a_t \mid h_t)$0, and $\pi(a_t \mid h_t)$1 define obligation, permission, and prohibition, while audit functions map traces to compliance reports. This gives AgREE an explicit semantics for cognition, cooperation, and governance that complements data-driven adaptation (Dignum et al., 21 Nov 2025).

3. Architectural patterns and reasoning loops

The most explicit AgREE architecture is the KGC framework for emerging entities. Given a query $\pi(a_t \mid h_t)$2, the agent decides whether to answer from internal knowledge or call tools, starts with a basic retriever using the Wikipedia API, escalates to an advanced retriever using a Google Search API when needed, chunks retrieved documents into 3-sentence segments, filters chunks by keyword overlap, re-ranks them with MS-MARCO-MiniLM, reflects on sufficiency, and synthesizes a ranked answer list before entity linking to canonical KG ids (Zhao et al., 6 Aug 2025). The framework is zero-training, implemented with LangGraph, uses DeepSeek-V3 as the backbone LLM, enforces a strict <answer>...</answer> format, requires at least 10 ranked candidates, and uses a maximum of $\pi(a_t \mid h_t)$3 iterations, with average $\pi(a_t \mid h_t)$4 on Wikidata5M (Zhao et al., 6 Aug 2025).

The broader architectural literature generalizes this loop into production-grade LLM agents. “From Prompt-Response to Goal-Directed Systems” describes a reference architecture in which cognition is separated from control flow, memory, tool execution, governance, and observability. The architecture comprises an Agent Core, Control Layer, Memory Layer, Tooling Layer, Governance & Observability, and External Environment, with side-effecting actions mediated by typed, versioned tool interfaces, schema validation, sandboxes, RBAC, audit logs, tracing, and budgeted autonomy (Alenezi, 11 Feb 2026). Its generic loop constructs context from state, memory, and policies; proposes a plan step; repairs plans that violate policy; executes typed tool calls; updates state and memory; and terminates through summarize-or-fail-safe conditions (Alenezi, 11 Feb 2026).

This architectural pattern reappears in domain-specific AgREE mappings. The survey on agentic reasoning treats ReAct, Tree-of-Thoughts, Graph-of-Thought, Plan-and-Execute, and KG-RAG as orchestration patterns for open-world discovery and action, while memory systems such as MemGPT, MEM0, Zep, Workflow Memory, and graph-structured memory are presented as persistent substrates for representing novelty across time (Wei et al., 18 Jan 2026). This suggests that AgREE is less a single control loop than a class of loops in which retrieval, planning, memory writing, self-reflection, and policy enforcement are first-class operations.

4. Collective reasoning, belief alignment, and communication

AgREE becomes explicitly collective when coherent behavior depends on aligning evolving internal belief states across agents. In 6G, “Reasoning-Native Agentic Communication for 6G” defines belief divergence as the case where two agents correctly interpret the same semantic information yet still generate inconsistent actions because their internal reasoning processes evolve differently (Seo et al., 19 Feb 2026). The proposed reasoning-native architecture augments the conventional stack with a Reasoning Coordination Plane, grounded in a shared ontology and a Recursive Belief Engine (RBE), that structures ontology-indexed belief variables, predicts counterpart reasoning under bounded belief modeling, estimates misalignment, and triggers communication when predicted divergence crosses a threshold. The paper also treats silence as information, uses alignment tokens and latent-space anchors for periodic resynchronization, and invokes OOD detection and hard resets when anomalies are detected (Seo et al., 19 Feb 2026).

The reported evaluations make the coordination claim concrete. RAS rises from 42% for Classical and 68% for Semantic to 94% for Agentic (MAR); DIB is 3.5× higher than Semantic; MBS falls to 60% over 1000 cycles under semantic drift but Agentic maintains over 91%; and in a reported overhead comparison Classical has 180% overhead relative to the semantic baseline 100%, whereas Agentic has 58% overhead with the highest success rate of 92% (Seo et al., 19 Feb 2026). In this formulation, communication is not merely transmission of bits or meanings; it is a regulator of distributed reasoning.

A related but distinct multi-agent formulation appears in SALT, a framework for uncertain emerging events in streaming environments. SALT uses specialized agents with topic sets, a dynamic topology $\pi(a_t \mid h_t)$5, and belief tuples $\pi(a_t \mid h_t)$6 containing statement text, confidence, timestamp, and references. Connection weights follow $\pi(a_t \mid h_t)$7, and lateral information flow occurs through selective propagation of relevant beliefs over long-distance agent interactions (Dernbach et al., 2024). On the reported benchmarks, Temporal-Claude-Sonnet-3.5 achieves RP 45.46% and HQ 26.88%, while SaLT-Claude-Sonnet-3.5 reaches RP 63.54% and HQ 34.68%; Temporal-GPT-4o reaches RP 35.83% and HQ 18.28%, while SaLT-GPT-4o reaches RP 57.33% and HQ 34.30% (Dernbach et al., 2024). Across both 6G and SALT, the recurring AgREE theme is that collective reasoning quality depends on belief management, structured communication, and topology control rather than on isolated model competence.

5. Benchmarks, domains, and empirical results

The original AgREE benchmark contribution is the Emerging-Entities test-only split for KGC, constructed from Wikidata with a time range from January 1, 2025 to May 1, 2025, containing 3,664 test triples, 90 relations, 332 emerging entities, and 20 categories (Zhao et al., 6 Aug 2025). On this dataset, AgREE reports Hits@1 45.2%, Hits@5 60.3%, Hits@10 75.5%, and MRR 55.3, compared with ReAct at 30.2% on Hits@1/5/10 and MRR 30.2, SimKGC at Hits@1 14.8%, Hits@5 18.7%, Hits@10 20.4%, MRR 16.9, and KGT5-Context+Desc at 0.0% across metrics (Zhao et al., 6 Aug 2025). The paper also introduces relation-aware Hits@N to address relations with multiple valid targets, arguing that standard Hits@N can be unfair in such settings (Zhao et al., 6 Aug 2025).

GeoBrowse extends the AgREE problem into multimodal, evidence-grounded geolocation. The benchmark contains 300 instances, with 199 in Level 1 and 101 in Level 2, and combines weak visual cue composition with BrowseComp-style multi-hop verification under entity obfuscation (Geng et al., 5 Apr 2026). Its agentic workflow, GATE, uses five think-with-image tools—Crop, Rotate, Auxiliary Lines, Local Super-Resolution, and Pixel Analysis—and four knowledge-intensive tools—Web Image Search, Web Text Search, Visit, and Code Interpreter (Geng et al., 5 Apr 2026). The paper reports that GATE improves GPT-4o from 23.1% to 31.8% on Level 1 and from 11.9% to 21.8% on Level 2, improves GPT-5 from 27.6% to 39.7% on Level 1 and from 17.8% to 30.7% on Level 2, and that Gemini-3-Pro achieves the best overall scores of 48.2% on Level 1 and 34.7% on Level 2 (Geng et al., 5 Apr 2026). A central finding is that gains come from coherent, level-specific tool-use plans rather than more tool calls.

Enterprise AgREE mappings emphasize auditable multi-hop reasoning over heterogeneous systems. HEAR, built on a Stratified Hypergraph Ontology with a base Graph Layer and a Hyperedge Layer, runs an evidence-driven reasoning loop over ERP, SRM, WMS, and BPM interfaces without retraining the LLM (Wang et al., 14 May 2026). On order-fulfillment blockage root cause analysis, HEAR Complete reports 94.7% accuracy with 6.3±0.2 turns and 38.1K±1.3 tokens, compared with 74.2% for HEAR Declarative, 52.6% for Table-RAG, and 11.3% for CSR-RAG (Wang et al., 14 May 2026). The same work reports that Qwen-3.5-27B reaches 92.7% on OF-BRCA, indicating that ontology grounding and tool execution can match proprietary-level performance with open-weight models (Wang et al., 14 May 2026).

6. Governance, risk, and unresolved problems

A persistent claim across the literature is that agentic reasoning about emerging entities is inseparable from runtime governance. The systems-theoretic account identifies deceptive alignment (alignment faking), self-exfiltration, sandbagging, disabling oversight, self-deception, subgoal-driven autonomy, human agency loss, multi-agent competition and coordination failures, and the accumulation of local low-risk decisions into systemic hazards as system-level risks (Miehling et al., 28 Feb 2025). Its proposed mitigations include residual control rights, human approval gates, subgoal monitoring, structured uncertainty reporting, and governance of human–agent communication (Miehling et al., 28 Feb 2025).

The institutional turn makes these constraints explicit. “Agentic AI, Retrieval-Augmented Generation, and the Institutional Turn” introduces a governance graph

$\pi(a_t \mid h_t)$8

where states carry capability restrictions $\pi(a_t \mid h_t)$9, and a sanction function

$M = \langle U, V, F, P(U) \rangle$0

with the explicit condition that when sanctions exceed the gains from deviation, compliant behavior becomes the dominant strategy (Osmond, 20 Feb 2026). Here RAG is treated as externalized epistemic infrastructure: outputs are grounded in verifiable repositories and produce documentary audit trails instead of opaque parametric emissions (Osmond, 20 Feb 2026). The AAMAS-oriented account complements this with norms, institutions, role assignments, social commitments, and audit semantics over trace logs (Dignum et al., 21 Nov 2025).

At the implementation level, governance appears as typed tool contracts, policy engines, short-lived credentials, least privilege, end-to-end structured tracing, immutable audit logging, circuit breakers, and high-risk human approvals (Alenezi, 11 Feb 2026). STAR-XAI pushes this further into interactive protocol design through a Consciousness Transfer Package (CTP), a Gameplay Cycle with ante-hoc justification, an Absolute Verification Module (AVM), Proposal Synchronization Protocol (PSP), and state-locking checksums that set a state pointer to the last validated state and are reported to eradicate state-based hallucinations in the case study (Guasch et al., 22 Sep 2025). This suggests that one branch of AgREE research is converging on explicit governance artifacts—rulebooks, checksums, sanctions, traces, and approval gates—as integral components of reasoning systems rather than external compliance wrappers.

Open problems are correspondingly broad. The literature highlights the balance between pretraining and environment-driven exploration, trust transfer across task families, scalable monitoring of subgoal generation, human–agent communication scaffolds, pairwise RBE scaling as $M = \langle U, V, F, P(U) \rangle$1, verifiability, interoperability, safe autonomy, world modeling, long-horizon interaction, scalable multi-agent training, and governance for real-world deployment (Miehling et al., 28 Feb 2025). Across these accounts, the unresolved question is not whether agents can reason about emerging entities, but how such reasoning can remain current, coherent, auditable, and controllable as entities, tools, institutions, and interaction structures continue to change (Wei et al., 18 Jan 2026).