Autonomous Industrial Control using an Agentic Framework with Large Language Models

As chemical plants evolve towards full autonomy, the need for effective fault handling and control in dynamic, unpredictable environments becomes increasingly critical. This paper proposes an innovative approach to industrial automation, introducing validation and reprompting architectures utilizing large language model (LLM)-based autonomous control agents. The proposed agentic system, comprising of operator, validator, and reprompter agents, enables autonomous management of control tasks, adapting to unforeseen disturbances without human intervention. By utilizing validation and reprompting architectures, the framework allows agents to recover from errors and continuously improve decision-making in real-time industrial scenarios. We hypothesize that this mechanism will enhance performance and reliability across a variety of LLMs, offering a path toward fully autonomous systems capable of handling unexpected challenges, paving the way for robust, adaptive control in complex industrial environments. To demonstrate the concept’s effectiveness, we created a simple case study involving a temperature control experiment embedded on a microcontroller device, validating the proposed approach.

💡 Research Summary

The paper addresses a critical gap in modern industrial automation: the inability of conventional rule‑based or reinforcement‑learning controllers to cope with rare, unforeseen disturbances (“long‑tail” events) that frequently occur in chemical plants and other complex processes. To overcome this limitation, the authors propose a novel agentic framework that places a large language model (LLM) at the core of a multi‑agent control loop. The system consists of four specialized agents—Monitoring, Actor, Validator, and Reprompter—interacting through a digital‑twin simulation that serves as a safe sandbox for action testing.

The Monitoring Agent continuously ingests sensor data and decides whether an anomaly has occurred; if so, it triggers the downstream agents. The Actor Agent uses the LLM to generate a concrete control command (e.g., “turn heater on/off”) based on the current goal. Before any command reaches the physical plant, it is fed to a digital twin that predicts the outcome of the action without risking equipment. The Validator Agent then checks the simulated outcome against predefined safety and performance criteria. When the action passes validation, it is deployed to the real system; otherwise, the Validator flags the failure.

At this point the Reprompter Agent enters the loop. It takes the feedback from the digital twin and the Validator, reformulates the prompt, and asks the Actor (LLM) to produce a revised action. This “generate‑validate‑re‑prompt” cycle repeats until either the action satisfies the validator or a maximum number of iterations is reached. By explicitly separating the LLM’s creative reasoning from a deterministic safety check, the architecture mitigates the well‑known hallucination problem of generative models, providing a practical safety net for high‑risk industrial settings.

To demonstrate feasibility, the authors implement the framework on an Arduino‑based TCLab temperature‑control platform. The task is a simple hysteresis control: the heater must be turned off when temperature exceeds 27 °C and turned on when it falls below 25 °C. Over a 40‑minute run, the LLM‑driven Actor proposes heater actions, the digital twin simulates the thermal response, and the Validator either approves or rejects the proposal. When rejected, the Reprompter supplies corrective prompts, leading the Actor to adjust its command. The experimental results show reduced overshoot and undershoot compared with a naïve rule‑based controller, and the system recovers quickly from sudden temperature spikes, illustrating the benefit of the validation‑reprompt loop.

Despite its promise, the study leaves several open questions. First, the fidelity of the digital twin is crucial; any mismatch between the simulated and real dynamics could cause unsafe actions to be mistakenly approved. Second, the latency and computational cost of repeatedly invoking an LLM in a real‑time control loop are not quantified, raising concerns for high‑frequency processes. Third, the safety criteria and reprompting policies must be handcrafted by domain experts, limiting the claim of fully zero‑shot autonomy. Finally, the evaluation is confined to a single‑variable temperature control scenario; scalability to multi‑variable, high‑dimensional processes, or safety‑critical plants remains to be demonstrated.

In summary, the paper introduces an innovative architecture that couples LLM reasoning with systematic validation and feedback, offering a pathway toward more adaptable and resilient autonomous industrial control. The validation‑reprompt mechanism constitutes a valuable design pattern for future LLM‑based control systems, provided that digital‑twin accuracy, latency, and domain‑specific safety specifications are rigorously addressed in subsequent work.