Large language models (LLMs) make it possible for AI agents to do more than work through fixed procedures: they can plan and act on their own. For smart manufacturing, this opens up new possibilities, but it also brings particular hurdles. Two recent arXiv papers show two sides of the same trend: a systematic overview of industrial agents and a study showing how the costs of coding agents can be reduced in everyday use.
What happened
An overview of LLM-driven industrial agents
According to the paper "Industrial Agentic AI: LLM-Driven New-Generation Agents for Smart Manufacturing" from the cs.AI category, general-purpose agents cannot simply be deployed in manufacturing. The authors name three reasons: dynamic environments, complex physical constraints, and very strict reliability requirements.
The paper presents itself as a systematic overview and proceeds step by step:
- First, the authors analyze five central challenges and explain what makes an industrial agent.
- Then they develop a unified taxonomy with application scenarios, forms of embodiment, and five capability levels.
- Next, they present a four-layer system architecture, including core modules and mechanisms for closed-loop task execution.
- Finally, they cover representative applications along the industrial lifecycle and future research directions.
The stated goal is a structured reference for the design, development, and application of industrial agents.
Lower token costs for coding agents
The second paper comes from cs.CL and is titled "TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents". It concerns AI-assisted software development, where progress in leading coding agents has significantly increased token costs. Tokens are the text building blocks by which language models measure their work.
According to the authors, the efficiency of such agents is usually measured on isolated tasks. Developers, however, use them to continuously build and maintain large projects. The researchers therefore tested sequences of realistic tasks and found recurring, environment-dependent token costs that existing methods barely reduced.
TokenEconomizer is an external controller. It decides
- which operational micro-skills are given to the agent,
- which permanently irrelevant tools are hidden,
- when to intervene in unproductive error loops.
Testing was carried out with Claude Code, Codex, and GitHub Copilot CLI, each with leading models:
| Test environment | Reduction in average token costs |
|---|---|
| Coding agents in software repositories | 18.4 to 30.9 percent |
| Sequential web development and security audits | 6.4 to 21.2 percent, depending on task structure |
According to the authors, task success remained comparable. The savings appeared in every tested repository and in every tested combination of agent and model.
Background
From a fixed workflow to an autonomous agent
Classical automation follows predefined sequences: if event A occurs, step B follows. This works as long as the environment is stable. An LLM-driven agent, by contrast, is given a goal, breaks it down into sub-steps, selects tools, acts, and adjusts its plan when something does not go as expected.
In manufacturing this is appealing, because something is constantly deviating from the plan. But it is also risky: a text assistant that makes a mistake can be corrected. A incorrectly controlled machine can produce scrap or endanger people. This is why the authors emphasize physical constraints and reliability.
Why a taxonomy helps
In a fast-moving field, terminology is often inconsistent. "Agent" can mean a chatbot with access to tools or a robot. A taxonomy brings order to this. The paper distinguishes by application scenarios, by forms of embodiment (the shape in which an agent appears in the world), and by five capability levels. Such level models help clarify expectations: what can a system do reliably, and what remains with the human?
Closed control loop and costs
"Closed" means that the agent does not just plan and act, but also checks the result and adjusts its approach. Which modules the four layers contain in detail can be found in the full text. In general: the more feedback a system has, the earlier errors can be detected.
The TokenEconomizer study adds a practical perspective to this. In continuous operation, what matters is how much effort each task causes. An agent that repeatedly loads unnecessary tools or gets caught in error loops wastes tokens. The authors speak of "black-box" cost control, because the controller intervenes from the outside and does not need to modify either the agent or the model.
- PAWM: World Model Warns LLM Agents of Pitfalls
- OR-License: How Good Are LLM Agents at Operations Research?
- Microsoft: AI Agents as an Insider Risk in Companies
- AI agents break into more than 100 websites
- ACAGS: Fraud Detection for Paying AI Agents
What this means
For developers of agents: The overview can serve as a framework for classifying your own projects: which capability level does the system correspond to, and which of the five challenges weighs most heavily in the specific application? The four-layer architecture is useful as a point of comparison, but it does not replace testing in the real environment.
For manufacturing companies and smaller businesses: The papers are not products and make no promises regarding readiness for deployment. They do show, however, which questions should be clarified before a pilot project: How reliable must the system be? Which physical limits apply? Where does human oversight remain?
For teams using coding agents: The results suggest that costs during ongoing operation can be influenced without noticeably harming success. The ranges come from the study's test environments. How strong the effect is in your own project presumably depends on the task structure, the agent, and the model. It makes sense to monitor token costs over longer task sequences and not just for individual tasks.
For interested readers: Both papers are preprints and, as a rule, have not yet been reviewed by independent reviewers. The core message nevertheless stands: agents are becoming more capable, but reliability, structure, and cost determine whether they are fit for everyday use.
Frequently asked questions
What is the survey article about?
The article gives a systematic overview of industrial agents based on large language models (LLMs) that are meant to be used in smart manufacturing.
Why can't general-purpose AI agents simply be deployed in manufacturing?
According to the authors, dynamic environments, complex physical constraints and very high reliability requirements make it difficult to deploy general-purpose agents directly.
What do the authors specifically propose?
They develop a unified taxonomy covering application scenarios, forms of embodiment and five capability levels. In addition, they present a four-layer system architecture for executing tasks in a closed loop.
What does TokenEconomizer do?
In the tests, the external controller reduced the average token costs of coding agents by 18.4 to 30.9 percent, with comparable task success. In web development and security audits, the reduction was 6.4 to 21.2 percent.
Who is the survey article useful for?
It serves as a structured reference for the design, development and application of industrial agents.
