LLM-Driven Multi-Agent Framework

• LLM-driven multi-agent framework is a computational approach that assigns specialized roles to agents, enhancing task decomposition and collaborative problem-solving.
• Architectural patterns such as linear pipelines, DAG orchestration, and dynamic routing enable efficient dependency management and adaptive task execution.
• The system integrates structured memory, sensor fusion, and verifier-based feedback to ensure robust performance and reliable error correction.

An LLM-driven multi-agent framework is a computational architecture in which multiple specialized agents, each driven by a LLM or a related multimodal model, cooperate through explicit orchestration, structured intermediate representations, and tool use to solve tasks that exceed the reliability or scope of a single prompt–response interaction. In the recent literature, such frameworks appear in robotic autonomy, autonomous driving, document understanding, network auto-configuration, scientific workflow automation, and formal reasoning, where they decompose a high-level objective into role-specific subtasks, preserve intermediate state in reports or memory stores, and couple generation with verification, simulation, retrieval, or execution (Chen et al., 9 May 2025, Hou et al., 4 May 2025, Yu et al., 10 Mar 2025).

1. Conceptual foundations

The core premise of LLM-driven multi-agent design is that complex tasks are better handled by role differentiation than by a monolithic agent. This principle is explicit in frameworks that assign planning, execution, critique, memory, or domain-specialist roles to separate agents. In autonomous driving, KoMA argues that single-step reactive decisions are myopic, one-shot multi-step plans can become invalid when scene conditions suddenly change, and single-agent setups ignore the multi-vehicle interactions and conflicts common in ramp merges and roundabouts; it therefore combines multi-agent interaction, Goal–Plan–Action reasoning, shared memory, and ranking-based reflection (Jiang et al., 2024). In industrial automation, BMW Agents similarly treats planning, orchestration, tool schemas, memory, and collaboration as distinct agent-engineering concerns, organized around a Plan–Execute–Verify loop (Crawford et al., 2024).

A more abstract formalization appears in SIMAS, which defines a multi-agent system as

$$\mathcal{M} = \langle \mathcal{A}, \mathcal{Q}, \mathcal{H}, T, \Pi, \mathcal{O} \rangle,$$

with agent set, question space, conversation history, rounds, response function, and output space explicitly separated (Li et al., 30 May 2026). That paper also frames multi-agent performance as a trade-off between collaborative synergy and coordination overhead,

$P(n) = P_0 + S(n) - O(n),$

which is significant because it rejects the common assumption that adding agents monotonically improves performance (Li et al., 30 May 2026).

Across the literature, the “agent” is therefore less a conversational persona than a bounded functional unit with defined inputs, outputs, and tool affordances. This is visible in robotics-specific decompositions such as Task Analyst, Robot Designer, and Reinforcement Learning Designer (Chen et al., 9 May 2025), in engineering decompositions such as planning, generative design, surrogate prediction, optimization, and physics validation (Du et al., 8 Apr 2026), and in autoformalization pipelines that separate generation, hard critique, soft critique, import retrieval, and refinement (Zhang et al., 10 Oct 2025). This suggests that the defining property of the paradigm is not multiplicity alone, but explicit modularization of reasoning and action.

2. Architectural patterns and coordination mechanisms

Several coordination patterns recur. A simple form is the linear planner–executor pipeline with structured handoffs. In “Multi-Agent Systems for Robotic Autonomy with LLMs,” the workflow is strictly linear—Task Analyst → Robot Designer → RL Designer → Code Extractor → Report Extractor—with sectioned reports functioning as blackboard memory (Chen et al., 9 May 2025). BMW Agents adopts a Coordinator that invokes a Planner, manages a Task Queue over a DAG, dispatches tasks to Executors, and invokes a Verifier for independent validation (Crawford et al., 2024).

A more general pattern is graph-based orchestration. DynTaskMAS represents work as a dynamic DAG

$G = (V, E, W),$

with edge weights

$W(u,v) = \alpha \cdot C(v) + \beta \cdot I(u,v),$

where $C$ estimates compute complexity and $I$ estimates context transfer cost (Yu et al., 10 Mar 2025). This design supports asynchronous parallel execution, dependency-aware scheduling, and adaptive reconfiguration. TurboAgent also uses LangGraph to construct a task graph with conditional branches, allowing the LLM planner to decide whether to generate, optimize, or validate based on intermediate results (Du et al., 8 Apr 2026).

Other frameworks emphasize dynamic routing rather than fixed flow. The document-understanding system in “Parallelism Meets Adaptiveness” builds a dependency graph, routes straightforward subtasks to single agents, triggers parallel competition for ambiguous subtasks, stores all candidates in shared memory, and lets an evaluator select the best output (Xia et al., 22 Jul 2025). EngiAI uses a hierarchical supervisor architecture in LangGraph, where a central supervisor classifies intent and routes requests to engineering, RAG, HPC, CLI, or printer-control agents (Molinari et al., 19 May 2026). M2HRI, by contrast, uses centralized coordination over socially differentiated robots: each agent maintains its own perception–memory–planning loop, while a coordinator computes response likelihoods and selects which robots should speak sequentially (Hasan et al., 13 Apr 2026).

Pattern	Characteristic mechanism	Example papers
Linear pipeline	Sequential reports or artifacts passed downstream	(Chen et al., 9 May 2025, Crawford et al., 2024)
DAG orchestration	Dependency graph, weighted scheduling, parallel ready tasks	(Yu et al., 10 Mar 2025, Du et al., 8 Apr 2026)
Dynamic routing	Task reassignment, evaluator-based selection, feedback bus	(Xia et al., 22 Jul 2025, Molinari et al., 19 May 2026)
Centralized coordination	Global selector chooses responding agents or communication structure	(Hasan et al., 13 Apr 2026, Bae et al., 18 May 2026)

The architectural choice is usually domain-driven. Robotics and auto-configuration emphasize deterministic staging and verifier-controlled regeneration (Chen et al., 9 May 2025, Sarıdaş et al., 31 May 2026), while open-ended document understanding and HRI emphasize adaptiveness, confidence-aware routing, and interaction management (Xia et al., 22 Jul 2025, Hasan et al., 13 Apr 2026).

3. Memory, communication, and grounding

A central technical issue is how agents share state without collapsing into an unstructured long context. One widespread solution is structured memory. In the robotic autonomy framework, reports are constrained to specific sections and act as persistent memory/state (Chen et al., 9 May 2025). BMW Agents separates short memory for within-task message history from episodic memory stored in a vector database for indirect dependencies and experiential learning (Crawford et al., 2024). KoMA uses a shared Chroma vector database keyed by text-embeddings-ada-002 to accumulate scenarios, planning traces, actions, and evaluation scores across agents (Jiang et al., 2024).

DynTaskMAS extends this memory perspective into explicit systems design. Its Semantic-Aware Context Management System provides hierarchical distributed memory, semantic indexing, relevance-based dissemination, access control, and two-phase commit updates (Yu et al., 10 Mar 2025). M2HRI partitions memory into working memory, semantic memory, and episodic memory, with per-agent private stores implemented as ChromaDB collections and mediated by an LLM-based memory controller (Hasan et al., 13 Apr 2026). In MASA, state is passed as tuples such as $(s,\phi,c,e,d,j)$ and grounded in a knowledge base plus theorem-prover feedback (Zhang et al., 10 Oct 2025).

Grounding mechanisms vary with task structure. Retrieval-augmented grounding is explicit in Beyond-5G auto-configuration, where example configurations and base parameter metadata are indexed with FAISS and all-MiniLM-L6-v2 embeddings; generation is segmented, verifier-guided, and corrected through selective regeneration of invalid segments (Sarıdaş et al., 31 May 2026). DriveAgent instead grounds multimodal reasoning in time-synchronized camera, LiDAR, GPS, and IMU streams, with cross-sensor consistency computed through Euclidean distances such as

$\Delta_{i,j}(t) = \|v_i(t) - \ell_j(t)\|$

and diagnostic aggregation over structured sets and triplets (Hou et al., 4 May 2025). CellAgent grounds tool use through a curated tool registry, standardized docstrings, and a code sandbox, thereby reducing API hallucinations in scientific analysis (Xiao et al., 2024).

Communication is therefore not merely conversational. It includes sectioned reports, JSON outputs, message buses, vector-memory retrieval, structured deltas, and executable code. A plausible implication is that the most stable LLM-driven multi-agent systems are those that externalize coordination state into typed artifacts rather than relying on raw dialogue alone.

4. Domain-specific realizations

The framework family is notably cross-domain. In robotics, one line of work converts task descriptions into robot design and control artifacts. The robotic autonomy system produces technical analysis reports, path visualizations, and runnable RL code files—env.py, train.py, and eval.py—from a natural-language description specifying environment, base and target locations, and arm-length options (Chen et al., 9 May 2025). A related mechatronics framework uses a High-Level Planning Agent to coordinate Mechanical Design, Simulation and Validation, Electronics Design, and Embedded Software agents for an autonomous dual-propeller vessel, with human approval gates for turbulent CFD configuration and electronics safety (Wang et al., 20 Apr 2025). TurboAgent further integrates generative geometry creation, Transformer-based performance prediction, LLM-guided optimization, and CFD/FEA validation for a transonic compressor (Du et al., 8 Apr 2026).

In autonomous driving, the emphasis shifts from geometry synthesis to situational reasoning. DriveAgent decomposes perception and reasoning into Descriptive Analysis, Vehicle Reasoning, Environmental Reasoning, and Response Generation, with explicit sensor fusion over camera, LiDAR, GPS, and IMU (Hou et al., 4 May 2025). KoMA addresses multi-vehicle interaction through intention inference, Goal–Plan–Action planning, shared memory, and ranking-based reflection (Jiang et al., 2024). LMAC uses an LLM offline to synthesize executable communication code for cooperative MARL, so that agents exchange compact numeric messages designed to maximize state reconstruction accuracy and reduce inter-agent imbalance, without online LLM calls during execution (Bae et al., 18 May 2026).

In social and embodied interaction, M2HRI assigns each robot explicit Big Five personality traits, persistent long-term memory, and a centralized coordination mechanism conditioned on those differences (Hasan et al., 13 Apr 2026). In document-centric tasks, the adaptive framework for SEC 10-K analysis uses dynamic task routing, bidirectional feedback, and parallel agent evaluation to improve factual coverage, coherence, and efficiency (Xia et al., 22 Jul 2025). In engineering workflow automation, EngiAI coordinates topology optimization, document retrieval, SLURM job orchestration, and 3D printer control through a supervisor plus seven specialist agents (Molinari et al., 19 May 2026).

Scientific and formal domains exhibit the same architectural logic under different constraints. CellAgent automates scRNA-seq analysis through Planner, Executor, and Evaluator roles, hierarchical decision-making, and self-iterative optimization (Xiao et al., 2024). AstroReview automates telescope proposal assessment in three stages—novelty and scientific merit, feasibility and expected yield, and meta-review and reliability verification—with a Proposal Authoring Agent and a Decision Gate for iterative refinement (Wang et al., 31 Dec 2025). MASA coordinates LLMs, theorem provers, retrieval, and hard/soft critique agents to transform natural-language mathematics into Isabelle/HOL or Lean4 code (Zhang et al., 10 Oct 2025).

These instantiations differ in toolchains and artifacts, but the recurrent pattern is the same: a planner or coordinator decomposes the task; specialists operate over bounded representations; validators or evaluators close the loop.

5. Evaluation, scaling behavior, and empirical performance

Evaluation is highly domain-specific, but several recurring axes appear: task completion, artifact validity, alignment with requirements, efficiency, and robustness under ambiguity or scale. In robotic autonomy, experiments over ten industrial and medical scenarios use TCP, CEF, MA, RDA, and RM; the reported model ranking is GPT-4o-mini ≈ DeepSeek-V3 < GPT-4o < DeepSeek-R1, and the system can design feasible robots with control strategies when appropriate task inputs are provided (Chen et al., 9 May 2025). In autonomous driving, DriveAgent reports best object-detection Precision and F1 at 89.96 and 71.62, strong vision-reasoning accuracy under misaligned camera distractors, and environmental reasoning accuracy of 58.95 on R2 and 65.71 on R3 (Hou et al., 4 May 2025).

In workflow systems, DynTaskMAS reports execution time reductions of 21.3% for simple tasks, 27.6% for medium tasks, and 33.0% for complex tasks, with resource utilization increasing from 65% to 88% and throughput scaling to 42.7 tasks/s at 16 agents (Yu et al., 10 Mar 2025). EngiAI reports 96–97% average task completion for proprietary models on Beams2D, while open-source 4B models reach 55–78%; its RAG benchmark yields scores near 1.0 with retrieval enabled and near zero without retrieval by gate design (Molinari et al., 19 May 2026). In Beyond-5G auto-configuration, the decomposed generation and decomposed verification pipeline reaches 94.4% success and improves average success rate by 22.7% over monolithic methods (Sarıdaş et al., 31 May 2026).

Some papers directly compare multi-agent systems to strong baselines. KoMA improves ramp-merging success from 30% untrained to 70% after 40 episodes, versus 65% for a deep MARL baseline trained for 40k episodes, and reaches 80% success on a roundabout scenario with transferred memory (Jiang et al., 2024). Meta-Team reports the best performance across nine benchmark columns, with an average score of 62.7 and average gains of +6.3 points over MASFly and +6.6 over a hand-crafted MAS (Hao et al., 28 May 2026). AstroReview identifies genuinely accepted proposals with positive accuracy of 87.08% under the Multi-Reviewer + Meta-Review + Reliability configuration and raises acceptance probability of revised drafts from about 33% in Round 1 to 99% in Round 2 in its iterative refinement module (Wang et al., 31 Dec 2025).

A separate line of evaluation asks whether agent scaling itself helps. SIMAS shows that, for sufficiently capable base models, performance follows an inverted-U rather than monotonic improvement as agent count increases; for small models, performance often degrades monotonically with more agents (Li et al., 30 May 2026). The paper’s fixed-context ablation shows that the decline persists even when input length is equalized, indicating that the governing factor is coordination overhead rather than long-context failure alone (Li et al., 30 May 2026). This is one of the clearest empirical rebuttals to the notion that agent plurality automatically yields collective intelligence.

6. Reliability, misconceptions, and future directions

A persistent misconception is that LLM-driven multi-agent systems are inherently more robust than single-agent systems. The literature does not support that view. Several frameworks explicitly introduce verifiers, evaluators, or reflection loops because raw multi-agent interaction remains error-prone. In auto-configuration, hallucinated configuration parameters are detected by a Configuration Verification Agent and corrected through low-computational segment-level regeneration (Sarıdaş et al., 31 May 2026). AstroReview adds a Reliability Verifier to correct schema deviations, logical inconsistencies, and unsupported assertions (Wang et al., 31 Dec 2025). MASA uses theorem provers as hard critics, because autoformalization cannot rely on LLM semantic judgment alone (Zhang et al., 10 Oct 2025).

Another misconception is that communication quality follows automatically from having more messages. LMAC shows that communication must be designed against explicit state-awareness criteria, and its cycle-consistency term is introduced precisely to suppress non-reconstructable or redundant content (Bae et al., 18 May 2026). DriveAgent similarly relies on explicit cross-sensor consistency distances and structured outputs, rather than unconstrained fusion, to detect misalignment and noise (Hou et al., 4 May 2025). In document understanding, feedback loops and shared memory are not optional enhancements: removing shared memory or feedback caused more than 20% drops in coverage and coherence (Xia et al., 22 Jul 2025).

Common limitations are also remarkably consistent. Many systems remain sim-only or offline: the robotics pipeline has no CAD/URDF generation, no obstacle modeling, and no sim-to-real transfer (Chen et al., 9 May 2025); the mechatronics framework requires human intervention for turbulent CFD setup (Wang et al., 20 Apr 2025); TurboAgent depends on trained surrogates and domain-specific toolchains (Du et al., 8 Apr 2026). Several papers do not report latency or real-time constraints, including DriveAgent and M2HRI (Hou et al., 4 May 2025, Hasan et al., 13 Apr 2026). Dataset scope is often narrow: DriveAgent is restricted to three routes at one location (Hou et al., 4 May 2025), and AstroReview evaluates Stage 3 only because full proposal texts were unavailable (Wang et al., 31 Dec 2025).

The future directions proposed across papers are convergent. They include richer multimodality and perception, stronger verification and safety constraints, more adaptive workflow management, explicit conflict resolution, and better integration of external simulators or formal tools. Robotics work calls for obstacle and dynamic-object reasoning, URDF-based reachability, and stronger simulation checks (Chen et al., 9 May 2025). Workflow systems point toward RL or Bayesian optimization of orchestration policies (Yu et al., 10 Mar 2025). HRI work points toward larger teams, longer deployments, and more formal memory policies (Hasan et al., 13 Apr 2026). Evolutionary MAS work points toward theories of self-evolution, causal telemetry, and tighter provenance tracking (Hao et al., 28 May 2026).

Taken together, the literature presents the LLM-driven multi-agent framework not as a single architecture, but as a design space organized around decomposition, structured state exchange, grounding, verifier-mediated correction, and domain-specific orchestration. The dominant research question is no longer whether multiple LLM agents can cooperate, but under what representational, organizational, and evaluative constraints such cooperation remains auditable, efficient, and technically reliable.