Skip to content

Hybrid Agentic AI and Multi-Agent Systems in Smart Manufacturing

🕒 Published (v1): 2025-11-23 03:06 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper proposes a layered hybrid framework that combines LLM-based orchestration agents with rule-based and SLM-powered specialized agents for Prescriptive Maintenance (RxM) in smart manufacturing. The LLM orchestrator acts as a meta-controller—selecting tools, adapting strategies, and routing tasks across perception, preprocessing, analytics, and optimization layers. A proof-of-concept is validated on two industrial datasets.

Problem

Traditional Multi-Agent Systems (MAS) in manufacturing rely on predefined rules and static task-specific models, limiting adaptability to unforeseen events. Meanwhile, current agentic AI systems are mostly single-agent architectures that lack distributed scalability and the robustness required for industrial environments. No prior work integrates LLM-powered agentic reasoning with MAS coordination for manufacturing RxM.

Method

A five-layer hybrid architecture (intelligence, perception, preprocessing, analytics, optimization), each managed by specialized agents with structured input/output interfaces. The LLM Orchestration Agent (Gemini 2.5 Flash) maintains a workflow context object encoding goal, available tools, completed steps, performance metrics, and lessons learned, then generates validated instructions for downstream agents via structured prompts. Retry logic (up to three attempts) handles malformed outputs; fallback defaults to rule-based execution if the LLM is unavailable. SLM agents (Qwen3:4B via Ollama, run locally) handle lightweight tactical decisions (model selection, hyperparameter suggestions). The preprocessing agent uses a rule-based tool decider: e.g., KNNImputer (\(k=3\)) if missing values \(>20\%\), SimpleImputer (median) for \(10\text{–}20\%\), RobustScaler if dataset \(>100\) MB. The analytics agent supports classification, regression, and anomaly detection; an Adaptive Intelligence module triggers automated model search if \(R^2 < 0.1\) or accuracy \(< 0.6\). The optimization agent computes priority scores using

\[\text{Priority Score} = \frac{\text{Predicted value} - \text{mean}}{\text{std}}\]

with warning/critical thresholds at mean \(+2\sigma\) and mean \(+3\sigma\).

Key Contributions

  • Layered hybrid architecture combining LLM orchestration with SLM edge agents and rule-based tactical components for manufacturing RxM.
  • An LLM Orchestrator that maintains persistent workflow context (goals, tools, step history, lessons learned) and drives adaptive strategy selection across agents.
  • SLM integration (local, privacy-preserving) for low-latency tactical decisions without cloud dependency.
  • Automated schema discovery, preprocessing pipeline construction, and adaptive model selection within a multi-agent workflow.
  • Human-in-the-loop (HITL) controls at perception and approval stages for auditability.
  • Proof-of-concept implementation validated on two industrial manufacturing datasets (open-sourced on GitHub).

Results

  • The framework successfully auto-detected schemas and adapted preprocessing pipelines on both datasets without manual intervention.
  • Adaptive Intelligence module activated when baseline models fell below thresholds, identifying better-performing configurations automatically.
  • Actionable, prioritized maintenance recommendations with estimated cost and timeframes were generated.
  • No quantitative baseline comparison figures (accuracy/F1/R² numbers against named baselines) are reported in the available text—the evaluation is described as qualitative proof-of-concept validation.

Limitations

  • No quantitative comparison against baseline single-agent or classical MAS approaches; results are descriptive, not ablative.
  • Proof-of-concept uses offline datasets only; real-time IIoT streaming integration is deferred to future work.
  • LLM orchestrator (Gemini 2.5 Flash) relies on cloud inference, introducing latency and privacy concerns the paper acknowledges but does not fully resolve.
  • SLM reasoning capacity is acknowledged to be limited relative to LLMs; boundary conditions for escalation are rule-based, not learned.
  • Framework tested on only two datasets; generalizability to diverse manufacturing domains is unverified.

Relevance to Harnesses / Meta-Harnesses

This paper is a direct instantiation of the meta-harness pattern: the LLM Orchestration Agent functions as a programmable harness that sequences heterogeneous sub-agents, manages workflow state across steps, enforces retry/fallback logic, and adapts execution strategy based on accumulated context—exactly the orchestration layer a meta-harness provides. The explicit separation of orchestrator (LLM) from executors (SLM, rule-based agents) mirrors the harness/worker split common in harness architectures. The persistent workflow context object (goal, tools, step history, lessons learned) is a runtime harness state store, and the validated structured-output contract between orchestrator and downstream agents is analogous to harness API contracts. For someone tracking harness design, this paper offers a concrete applied instantiation of LLM-as-meta-harness with fallback, retry, and HITL hooks in a safety-critical domain.