Agentic Model: Bounded Autonomous Framework

• Agentic models are frameworks that exhibit bounded autonomy by integrating multi-step reasoning, tool use, memory, and multi-agent coordination.
• They implement diverse architectures—from hierarchical agent systems to probabilistic latent substructures—focusing on proactive execution rather than simple response generation.
• Applications span financial services, HCI, and organizational workflows, demonstrating enhanced task management, safety verification, and governance through structured autonomous operations.

An agentic model is a model or model-centered system that does more than generate outputs on request: it acts in a sustained, goal-directed manner, reasons over multi-step tasks, invokes tools and external systems, maintains context, and, in many formulations, coordinates with other agents or humans. The recent literature does not use the term univocally. Depending on the paper, it denotes an autonomous workflow participant in organizational operations, a coordinated multi-agent architecture with manager and worker roles, a formal host-agent/task-lifecycle abstraction, a human-centered policy model for deciding when and whether to act, or a latent probabilistic substructure inside a neural network. This suggests that “agentic model” is best understood as a family of models and model-centered frameworks for bounded autonomy rather than a single canonical architecture (Bandara et al., 27 Jan 2026, Acharya, 13 Mar 2026, Okpala et al., 8 Feb 2025, Allegrini et al., 15 Oct 2025, Jung et al., 26 Feb 2026, Lee et al., 8 Sep 2025).

1. Conceptual range and defining properties

Across the literature, the common denominator is a shift from passive response generation to execution under goals, constraints, and feedback. In organizational and enterprise papers, an agentic model is an autonomous system that can plan multi-step strategies, reason about goals, select and invoke tools, and execute workflows with limited human intervention; in workflow-centric formulations, the distinctive property is “agency in execution, not just intelligence in response” (Acharya, 13 Mar 2026, Bandara et al., 27 Jan 2026). In hierarchical multi-agent formulations, the term refers to a coordinated workflow built on top of an LLM, with manager agents delegating to specialized subordinate agents that use tools, memory, and role-specific prompts (Okpala et al., 8 Feb 2025). In human-centered HCI work, the agentic model is not primarily an architecture but a decision model for proactive intervention, built around Scene, Context, and Human Behavior Factors (Jung et al., 26 Feb 2026). In formal methods, the term is recast as a unified semantic object for orchestration and task management (Allegrini et al., 15 Oct 2025). In probabilistic theory, an “agent” is a probability distribution over outcomes with epistemic utility given by log score (Lee et al., 8 Sep 2025).

Formulation	Core idea	Representative paper
Workflow-centric organizational model	One or more AI agents operate as autonomous workflow participants under human supervision	(Bandara et al., 27 Jan 2026)
Enterprise autonomy model	Autonomous systems plan, reason, invoke tools, and execute business workflows	(Acharya, 13 Mar 2026)
Hierarchical crew model	Manager agents delegate to specialized worker agents with tools and memory	(Okpala et al., 8 Feb 2025)
Human-centered proactive model	Agent action is governed by Scene, Context, and Human Behavior Factors	(Jung et al., 26 Feb 2026)
Formal orchestration model	A host agent and task lifecycle define system behavior and verifiable properties	(Allegrini et al., 15 Oct 2025)
Latent probabilistic subagent model	Agents are outcome distributions combined through weighted logarithmic pooling	(Lee et al., 8 Sep 2025)

A recurrent misconception in this literature is that agentic models are simply “chatbots with tools.” Several papers reject that reduction explicitly. Agentic systems are described instead as systems of agents, tools, memory, and environment-mediated interaction; the relevant object is often the execution trace or workflow, not an isolated prompt-response pair (Wicaksono et al., 5 Sep 2025, Bandara et al., 27 Jan 2026). A second misconception is that an agentic model must be a single model architecture. The literature suggests the opposite: in many domains, the “model” is a socio-technical composite in which model inference, tool use, orchestration, governance, and human oversight are inseparable (Okpala et al., 8 Feb 2025, Park, 22 Feb 2026, Dignum et al., 21 Nov 2025).

2. Architectural patterns and capability structure

Several papers converge on a small set of capability families. A survey on future computing environments identifies four foundational patterns—Reflection, Tool Use, Planning, and Multi-Agent Collaboration—while enterprise and product-management papers repeatedly foreground autonomy, goal-driven behavior, dynamic task decomposition, persistent memory, and collaboration among specialized agents (Murad et al., 20 Sep 2025, Parikh, 1 Jul 2025, Bandara et al., 27 Jan 2026). AgentSeer’s decomposition of agentic systems into agents, tools, short-term memory, and long-term memory is a particularly concrete systems view, because it treats actions as individual LLM operations including response generation, tool calling, and agent communication, and components as the persistent entities through which execution unfolds (Wicaksono et al., 5 Sep 2025).

The architectural distinction between monolithic prompting and modular orchestration is especially sharp in finance and enterprise workflow papers. In the financial-services “crew” architecture, a “Data Science Manager” or “Model Risk Manager” routes work to specialized agents for exploratory data analysis, feature engineering, model selection, hyperparameter tuning, model training, evaluation, documentation compliance, model replication, conceptual soundness analysis, and outcome analysis; inter-agent communication is carried through a memory stream that records delegations, timestamps, tool inputs, action inputs, and context (Okpala et al., 8 Feb 2025). In organizational transition work, multi-agent coordination is favored because specialized agents with bounded roles improve modularity, maintainability, controllability, and parallelism, while humans move upward into supervisory and exception-handling roles (Bandara et al., 27 Jan 2026).

Multimodal papers extend this capability picture in a distinctive way. “Agentic-MME” frames multimodal agency as the combination of Visual Expansion and Knowledge Expansion: the model should actively manipulate images through cropping, rotating, enhancing, flipping, resizing, contrast adjustment, thresholding, denoising, or edge detection, and also retrieve missing knowledge through open-web search and page fetching (Wei et al., 3 Apr 2026). DeepEyesV2 operationalizes a similar view with code execution, image search, and text search, using a reasoning–tool–integration loop in which the model alternates among >, <code>, <tool_call>, and <answer> actions (Hong et al., 7 Nov 2025). This suggests that, in multimodal settings, agency is increasingly identified with the ability to acquire new evidence through interaction rather than to reason longer over fixed evidence.

3. Formal models and analytical lenses

The literature supplies several non-equivalent but complementary formalizations of agentic models. In attribution work, an LLM agent is modeled as a policy over interaction history,

$a_t = \pi_\theta(\cdot \mid H_t),$

with

$H_t = (o_0, a_0, o_1, a_1, ..., o_t),$

and the full trajectory transformed into temporally ordered components of types USER, THOUGHT, TOOL, OBS, and MEMORY (Qian et al., 21 Jan 2026). This is a trajectory-native view: the agent is defined by how actions arise from structured history, not by a single forward pass.

In formal verification work, the host agent itself is modeled as

$$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$

where $\mathcal{A}$ is the set of autonomous agents, $\mathcal{E}$ the set of external entities, $\mathcal{R}$ the registry, $\mathcal{C}$ the host-agent core, $\mathcal{O}$ the orchestrator, $\mathcal{C}_L$ the communication layer, and $\mathcal{S}_{\mathcal{H}}$ the host-agent state space (Allegrini et al., 15 Oct 2025). Subtasks are then given an explicit lifecycle

$H_t = (o_0, a_0, o_1, a_1, ..., o_t),$0

with states including $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$1, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$2, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$3, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$4, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$5, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$6, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$7, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$8, $H_t = (o_0, a_0, o_1, a_1, ..., o_t),$9, $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$0, and $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$1 (Allegrini et al., 15 Oct 2025). This formalization is notable because it makes multi-agent orchestration a state-transition system amenable to temporal-logic verification.

AAMAS-inspired work argues that contemporary agentic AI should be “agentified” through explicit cognitive structure, especially BDI architectures, communication protocols, mechanism design, trust models, and institutional modeling. In this formulation, beliefs $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$2, desires $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$3, and intentions $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$4 are not incidental implementation details but the explicit substrate of interpretable agency (Dignum et al., 21 Nov 2025). The paper’s deeper claim is that behaviorally autonomous systems are not yet fully agentic unless they also expose commitments, roles, norms, and coordination semantics.

Two additional mathematical lenses broaden the concept further. “On Agentic Behavioral Modeling” treats artificial agents as latent generative hypotheses about cognition and embeds them in joint probability models over task variables, latent agent variables, and observed actions (Ostwald et al., 30 Apr 2026). By contrast, “Probabilistic Modeling of Latent Agentic Substructures in Deep Neural Networks” identifies an agent with a probability distribution $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$5 over a finite outcome space $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$6 and epistemic utility $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$7, then defines compositions through weighted logarithmic pooling,

$$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$8

with $$\mathcal{H} = \left( \mathcal{A}, \mathcal{E}, \mathcal{T}, \mathcal{R}, \mathcal{C}, \mathcal{O}, \mathcal{C}_{L}, \mathcal{S}_{\mathcal{H}} \right),$$9 and $\mathcal{A}$0 (Lee et al., 8 Sep 2025). This is a markedly different notion of “agentic model”: agency becomes a probabilistic and compositional property of latent neural substructures rather than an explicit planner interacting with tools.

4. Governance, safety, and interpretability

Governance and assurance are central because agentic models are treated as actors that can continue operating after initial deployment. The most elaborate governance proposal is the Agentic AI Governance Maturity Model, a five-level framework spanning 12 governance domains and grounded in NIST AI RMF 1.0 and ISO/IEC 42001. Its simulation study runs $\mathcal{A}$1 scenario-level trials and reports large cross-level effects: from Level 1 to Level 5, the Sprawl Index falls from $\mathcal{A}$2 to $\mathcal{A}$3, Risk Incident Rate per 1,000 actions falls from $\mathcal{A}$4 to $\mathcal{A}$5, Effective Task Completion Rate rises from $\mathcal{A}$6 to $\mathcal{A}$7, Delegation Safety Rate rises from $\mathcal{A}$8 to $\mathcal{A}$9, and Net Business Value rises from $\mathcal{E}$0 to $\mathcal{E}$1. The same paper defines agent sprawl patterns—functional duplication, shadow agents, orphaned agents, permission creep, and unmonitored delegation chains—and argues that Level 3 is the minimum viable governance threshold (Acharya, 13 Mar 2026).

Interpretability work treats the question “why did the agent take this action?” as distinct from failure attribution. “The Why Behind the Action: Unveiling Internal Drivers via Agentic Attribution” formalizes completed trajectories into components and scores them through temporal likelihood dynamics,

$\mathcal{E}$2

followed by sentence-level perturbation analysis using probability drop, probability hold, and the combined score

$\mathcal{E}$3

Its default sentence scorer achieves Hit@1 $\mathcal{E}$4, Hit@3 $\mathcal{E}$5, and Hit@5 $\mathcal{E}$6 across nine trajectories (Qian et al., 21 Jan 2026). This work matters because it shifts agentic interpretability from “where did it fail?” to “which memory, tool result, or prior interaction drove this behavior?”

Safety verification is pushed further in “Agentic Model Checking,” which couples LLM agents with bounded model checking under the principle “agents propose, solvers verify.” Specifications are inferred top-down from caller context in a restricted DSL, verification is compositional at function granularity, and counterexamples are validated through reachability, callee feasibility, dynamic replay, and realism audit rather than being treated as immediate bug reports. The instantiated system, BMC-Agent, reports 62 confirmed real defects across LLM-generated kernel/compiler code and mature libraries, while also producing bounded clean verifications and selected functional-equivalence proofs (Sun et al., 20 May 2026).

Two additional papers expose risks that standard model evaluation misses. AgentSeer decomposes runs into action graphs and component graphs and shows systematic divergence between model-level and agentic-level vulnerability profiles: GPT-OSS-20B has model-level ASR $\mathcal{E}$7, Gemini-2.0-flash $\mathcal{E}$8, but agentic-level evaluation reveals “agentic-only” vulnerabilities, tool-calling with $\mathcal{E}$9 higher ASR, and agent transfer as the highest-risk tool category (Wicaksono et al., 5 Sep 2025). Agentic Problem Frames, by contrast, responds to open-loop failures with a closed-loop engineering pattern centered on dynamic specification and the Act-Verify-Refine loop,

$\mathcal{R}$0

treating verified outcomes as reusable knowledge assets and framing reliability as asymptotic convergence toward mission requirements $\mathcal{R}$1 (Park, 22 Feb 2026).

5. Human-centered and socio-technical interpretations

A substantial strand of the literature shifts the discussion from internal autonomy to human meaning, organizational adaptation, and labor relations. In “When Should an AI Act?”, the agentic model is a human-centered model for proactive action based on Scene, Context, and Human Behavior Factors. Scene contains Actor(s), Object(s), Background, and Activity; Context comprises spatial, temporal, interoceptive, individual, and social/cultural dimensions; Human Behavior Factors include attitude, perceived opportunity, perceived capability, motivation, and trigger. The paper’s design principles—behavioral alignment, contextual sensitivity, temporal appropriateness, motivational calibration, and agency preservation—make the agentic model a policy lens for deciding whether intervention should occur at all (Jung et al., 26 Feb 2026).

In product management, the PM-AI co-evolutionary model redefines agentic AI as a multi-agent system characterized by autonomy, goal-driven behavior, persistent memory, dynamic task decomposition, and multi-agent collaboration. Product managers are recast as orchestrators of socio-technical ecosystems operating through an “Adaptive Learning & Delegation Loop,” while governance remains centered on explainability, shared mental models, accountability, and strategic alignment (Parikh, 1 Jul 2025). A related systems survey argues that agentic AI may alter computing architecture itself, shifting workloads from massive public cloud services toward hybrid cloud-edge-on-premises environments because of resource efficiency, localized processing, and diminished data consumption footprints (Murad et al., 20 Sep 2025).

Other papers emphasize the human role in oversight rather than replacement. The organizational transition framework proposes a human-in-the-loop operating model in which humans act as orchestrators of multiple AI agents, while small, cross-functional, AI-augmented teams collaborate closely with business stakeholders to encode tacit domain knowledge into workflows (Bandara et al., 27 Jan 2026). By contrast, “The Shadow Boss” develops a sharply critical socio-technical interpretation through “Agentic Employment,” in which autonomous AI agents act as economic principals that directly hire, instruct, and pay human workers through XR-mediated micro-instructions. Its scenario construction identifies a liability void, atomized manipulation, cognitive deskilling, Diminished Reality, civic and social manipulation, and embodied extraction, characterizing workers as “biological actuators” or “human hardware” for invisible software entities (Lee, 14 Feb 2026). This suggests that the meaning of “agentic model” depends not only on tool use and planning but also on the institutional position from which autonomy is exercised.

6. Applications, benchmarks, and unresolved questions

The literature is already application-rich. In financial services, a dual-crew architecture uses GPT-3.5 Turbo and CrewAI to separate production modeling from model risk management. The modeling crew handles exploratory data analysis, feature engineering, model selection, hyperparameter tuning, training, evaluation, and documentation; the MRM crew performs documentation compliance checking, model replication, conceptual soundness review, stress testing, and documentation writing. Across fraud detection, credit approval, and portfolio credit risk datasets, the system is presented as a governed workflow rather than a single model call (Okpala et al., 8 Feb 2025). In organizations, tourism and transport workflows are decomposed into email-reading, filtering, availability, synthesis, planning-sheet, and publishing agents, exposed through MCP servers and invoked through LM Studio (Bandara et al., 27 Jan 2026).

Evaluation work shows both progress and immaturity. “Agentic-MME” introduces a process-verified benchmark with 418 real-world tasks, 6 domains, 3 difficulty levels, and over 2,000 stepwise checkpoints averaging 10+ person-hours of annotation per task; the best reported model, Gemini3-pro, reaches 56.3% overall accuracy, with a large drop on the hardest tasks (Wei et al., 3 Apr 2026). DeepEyesV2 introduces RealX-Bench with 300 QA pairs across 5 domains and reports 28.3 average accuracy for DeepEyesV2, with 19.5 on Perception, 22.5 on Reasoning, 28.9 on Search, and 18.1 on Integration (Hong et al., 7 Nov 2025). These results indicate that multimodal agency remains far from human performance even when code execution and web search are integrated.

Cost-aware deployment has become an application domain in its own right. Switchcraft is an inline router for agentic tool calling trained on 122,267 deduplicated function-calling examples spanning 14 categories and a pool of 8 candidate models. Its DistilBERT router achieves 82.94% accuracy at $\mathcal{R}$2 dollars per query, compared with 82.29% for the best individual model, GPT-5.3-chat, at $\mathcal{R}$3 dollars per query, amounting to about 84% lower cost and more than \$3,600 saved per million queries (Agarwal et al., 8 May 2026). This suggests that, in mature agentic systems, “the model” may increasingly denote a routed portfolio rather than a single fixed backbone.

Open problems are correspondingly heterogeneous. Formal frameworks remain split across symbolic orchestration models, probabilistic latent-agent models, cognitive-behavioral generative models, and human-centered policy lenses (Allegrini et al., 15 Oct 2025, Lee et al., 8 Sep 2025, Ostwald et al., 30 Apr 2026, Jung et al., 26 Feb 2026). Governance evidence is still often simulation-based rather than deployment-based (Acharya, 13 Mar 2026). Vulnerability evaluation remains deployment-specific, with model-level safety tests systematically missing agentic-only risks (Wicaksono et al., 5 Sep 2025). Multimodal benchmarks show that today’s systems often fail because they do not act, act unfaithfully, search poorly, or overthink (Wei et al., 3 Apr 2026, Hong et al., 7 Nov 2025). A plausible implication is that the next phase of research will be less about proving that models can call tools and more about reconciling five competing requirements: explicit cognitive structure, bounded autonomy, grounded verification, human agency preservation, and economically viable deployment (Dignum et al., 21 Nov 2025, Park, 22 Feb 2026, Agarwal et al., 8 May 2026).