Unified Human-Agent Collaboration Architecture

• Unified layered architecture is a systematic framework that formalizes human-agent collaboration by segregating processes into intent, orchestration, execution, and supervisory layers.
• It employs modular design principles such as explicit role separation and dynamic replanning to improve reliability, scalability, and error recovery in multi-agent systems.
• State-of-the-art implementations demonstrate adaptive scheduling, transparent oversight, and automated verification that enhance human supervision and optimize task assignment.

A unified layered architecture for human-agent collaboration formalizes the interaction, coordination, and oversight mechanisms required for intelligent, goal-driven teamwork between humans and multi-agent AI systems. Unified architectures address the complexity, cognitive limitations, and alignment challenges that arise when orchestrating heterogenous groups of agents (often LLM–based) alongside human supervisors, by enforcing explicit separation of concerns, transparent process control, and dynamic role/adaptation protocols. These architectures, as represented by state-of-the-art frameworks such as OrchVis, HA², HAWK, HADA, HMS-HI, and associated theory, yield scalable, inspectable, and adaptable collaboration pipelines.

1. Layered Decomposition: Theoretical Foundations

Unified layered architectures impose strict stratification, modularizing human-agent interaction into separable, composable subsystems. While the number and precise function of layers vary by system, several canonical divisions recur:

• Intent and Interaction Layer: Processes raw human instructions, parses intents, encodes constraints, and exposes surface-level interfaces for correction, mixed-initiative control, and feedback (Zhou, 28 Oct 2025, Wang et al., 13 Jun 2025, An et al., 9 Oct 2025).
• Process and Orchestration Layer: Encodes persistent, inspectable collaborative process models—e.g., goal hierarchies, workflow/task graphs, interdependencies, and allocation policies. This is the locus of structured goal decomposition, dynamic scheduling, and adaptation logic (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025, Wang et al., 13 Jun 2025).
• Execution and Verification Layer: Oversees agent execution, evaluation, low-level tool invocation, evidence collection, and machine-checkable verification against goal predicates (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025, Lamon et al., 2023).
• Conflict Management and Replanning Layer: Detects constraint violations, presents natural-language summaries and repair options, and supports partial or selective replanning for failed workflow branches (Zhou, 28 Oct 2025).
• Supervisory and Interactive Visualization Layer: Provides an interactive planning environment for visualization of dependencies, acceptance or override of assignments, and detailed inspection at any abstraction level (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025, Scibelli et al., 12 Dec 2025).
• Resource and Infrastructure Layers: Abstracts concrete models, databases, devices, and service endpoints, managing access, orchestration, and cross-layer protocol conversion (Cheng et al., 5 Jul 2025, Wang et al., 13 Jun 2025, An et al., 9 Oct 2025).

Formal models employ state triples $S^{(t)} = (I^{(t)}, P^{(t)}, F^{(t)})$, Petri net communication spaces, or hierarchical MDPs, ensuring lossless, compositional information flow, and support both MAS autonomy and deeply fused centaurian (hybrid human-AI) systems (Wang et al., 13 Jun 2025, Borghoff et al., 19 Feb 2025, Aroca-Ouellette et al., 7 May 2025).

2. Goal Parsing, Orchestration, and Assignment

Layered architectures begin with structured parsing of the user's free-form intent, leveraging LLMs with grammar and ontology constraints to synthesize a directed goal graph $G = \{g_1, ..., g_n\}$ whose nodes are coupled to domain ontologies and constraints. Each goal is supplemented with logical predicates for automated verification (e.g., $price(g) \leq 400$, $departureCity(g) = ATL$) (Zhou, 28 Oct 2025).

The orchestration layer then assigns subgoals to specialized agents using deterministic capability matching. For each required features vector $r_i$ and agent capabilities vector $s_j$, agent $a_j$ is chosen to maximize

$$\mathrm{match}(g_i, a_j) = r_i \cdot s_j - \alpha \|r_i - s_j\|$$

with $\alpha$ governing penalty for extraneous or missing skills. This produces a task graph $P = (V, E)$ representing the decomposed, executable plan (Zhou, 28 Oct 2025).

Other systems (e.g., HA²) implement a two-level MDP-based decomposition (manager: high-level task selection; worker: primitive action control), rigorously reflecting the distinction between sub-task planning and execution, with shared abstractions $\varphi(s)$ for agent-human policy alignment (Aroca-Ouellette et al., 7 May 2025).

3. Verification, Progress Tracking, and Adaptation

Execution monitoring implements automated verification of agent outputs against specified goal predicates, evaluating both hard and soft constraint satisfaction. The satisfaction score for goal $g_i$ is:

$$S_i = \frac{ \#\text{hard}_i^{\text{satisfied}} }{ \#\text{hard}_i^{\text{total}} } + \lambda \frac{ \#\text{soft}_i^{\text{satisfied}} }{ \#\text{soft}_i^{\text{total}} }$$

with $\lambda$ weighting soft constraints. Outputs failing to meet $S_i=1$ are flagged as incomplete or conflicting (Zhou, 28 Oct 2025).

Conflict detection mechanisms summarize violations in natural language, prompting the system to propose $k$ candidate repairs, which, after human inspection, are reincorporated as partial replans. Only affected subtrees of the workflow are updated, preserving maximal progress elsewhere. Selective, minimal replanning is a critical design feature, reinforcing human oversight without imposing full reruns (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025).

Adaptive scheduling modules (e.g., in HAWK) optimize task-to-agent assignments by maximizing resource-weighted throughput under real-time capacity and priority constraints, updating allocations as latency and failure reports are ingested (Cheng et al., 5 Jul 2025).

4. Human Oversight, Interactive Visualization, and Mixed Initiative

Unified architectures emphasize transparency and oversight through multi-panel dashboards. Top-level goal trees and task graphs display progress states, dependencies, and conflict indicators, while detailed summary panes expose rationales, predicted outcome metrics (cost, risk, completion time), and available repair actions. Manual overrides and autonomy-level controls ensure that the user can intervene or adjust at any abstraction layer, preserving strategic authority (Zhou, 28 Oct 2025, Scibelli et al., 12 Dec 2025).

Mixed-initiative collaboration is achieved via tightly-coupled interaction and process layers. Human input (edits, accepts, rejects) flows upward to update goals and allocations; system responses (proposals, status, repair candidates) are visualized or surfaced as actionable options. The human is kept strategically in-the-loop, but operational burden is minimized via orchestration (Zhou, 28 Oct 2025, Wang et al., 13 Jun 2025, Scibelli et al., 12 Dec 2025).

TIP (Time, Interaction, Performance) theory and associated agentic design heuristics—trust calibration, clarity, traceability—are explicitly realized through surface artifacts, schema-enforced interfaces, behavioral proxy agents, and orchestrator modules, yielding modular, auditable, and robust group modes from inception to execution (Scibelli et al., 12 Dec 2025).

5. Applications, Scalability, and Evaluation

Unified layered architectures have been validated across diverse domains:

• Travel Planning and Workflow Automation: OrchVis demonstrates multi-step LLM-agent collaboration and partial replanning with interactive oversight, showing qualitative human-centered scalability ($O(k)$ user effort vs. constant) (Zhou, 28 Oct 2025).
• Human-AI Socio-technical Teams: HMS-HI, in an urban emergency response simulation, yields 72% casualty reduction and 70% drop in cognitive load over baseline HiTL paradigms. Contributions of each layer (Shared Cognitive Space, DRTA, CSTC) are quantitatively ablated (Melih et al., 28 Oct 2025).
• Enterprise Decision Alignment: HADA formalizes cross-layer propagation of OKRs, KPIs, and ethical constraints with protocol-agnostic stakeholder agents, enabling fast remediation (e.g., ZIP code bias) and full decision lineage at scale (Pitkäranta et al., 1 Jun 2025).
• Multi-agent Creative Workflows: HAWK, via its adaptive scheduler and resource abstraction, demonstrated >92% agent invocation success and near-linear scaling to hundreds of agents and tasks without degradation (Cheng et al., 5 Jul 2025).
• Real-time Human-AI Collaboration: DPT-Agent combines FSM-based System 1 (low latency) with LLM-based System 2 (deliberative ToM, reflection), outperforming mainstream LLM frameworks in both efficiency and subjective human assessments (Zhang et al., 17 Feb 2025).

Coordination, knowledge extraction, protocol conversion, and experience sharing mechanisms (e.g., Co-TAP MEK layer) have enabled cross-agent compositionality, real-time event streaming, and cognitive chain distillation across application areas (An et al., 9 Oct 2025).

6. Design Principles, Limitations, and Research Directions

Key design principles derived from current unified architectures include:

• Orchestrate rather than micromanage—abstractions hide $O(k)$ operational details, exposing only strategic decisions (Zhou, 28 Oct 2025).
• Align high-level goals prior to task assignment—goal-first separation minimizes user confusion and workflow incoherence (Zhou, 28 Oct 2025, Scibelli et al., 12 Dec 2025).
• Explicitly model and visualize inter-agent dependencies and state—dependency surfaces and causal links yield greater user insight and error recoverability (Zhou, 28 Oct 2025, Scibelli et al., 12 Dec 2025).
• Support partial, non-blocking replanning and flexible adaptation—conflict repair is localized, yielding resilience under partial failure or new constraints (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025).
• Automate verification and maintain traceable, machine-checkable artifacts—predicate logic and state logs ensure inspectability and ex-post auditability (Zhou, 28 Oct 2025, Pitkäranta et al., 1 Jun 2025, Scibelli et al., 12 Dec 2025).
• Integrate dynamic protocol and resource adaptation for cross-domain and scalability needs (e.g., plug-and-play resource adapters, registry/discovery protocols) (Cheng et al., 5 Jul 2025, An et al., 9 Oct 2025, Wang et al., 13 Jun 2025).

Limitations noted in current research include:

• Absence of formal user studies or quantitative end-user results in some deployed systems, though qualitative HCI literature is cited (Zhou, 28 Oct 2025).
• The need for richer co-evolutionary infrastructure (dynamic memory, meta-reflection, process consistency audits) as workflows and team compositions scale (Wang et al., 13 Jun 2025).
• Remaining challenges in dynamic role/initiative management, process-awareness in interface design, and "undo" or rollback support in evolving workflow graphs (Wang et al., 13 Jun 2025, Scibelli et al., 12 Dec 2025).
• The dependency on LLM capabilities and prompt engineering for goal parsing and repair quality; failure cases can reduce subjective trust or require additional human intervention (Zhang et al., 17 Feb 2025).

Ongoing research directions seek enhanced process-aware interfaces, fine-grained performance tuning (potentially with reinforcement learning at the scheduling layer), domain-specific plugin packs, more robust cross-domain adaptability, and the formalization of cybernetic feedback for continual learning and adaptation (Cheng et al., 5 Jul 2025, Wu et al., 24 Apr 2025, An et al., 9 Oct 2025).

---

Unified layered architectures for human-agent collaboration represent the current synthesis of rigorous system design, formal process modeling, and adaptive, transparent workflow orchestration. Their adoption enables both large-scale agentic coordination and fine-grained, inspectable oversight by human operators, opening new domains for scalable, longitudinal, and ethically aligned intelligent collaboration (Zhou, 28 Oct 2025, Cheng et al., 5 Jul 2025, Melih et al., 28 Oct 2025, Wang et al., 13 Jun 2025, Aroca-Ouellette et al., 7 May 2025, Pitkäranta et al., 1 Jun 2025, Scibelli et al., 12 Dec 2025, An et al., 9 Oct 2025, Lamon et al., 2023, Wu et al., 24 Apr 2025, Borghoff et al., 19 Feb 2025, Zhang et al., 17 Feb 2025, Xin et al., 2024, Diggelen et al., 2019).