Table of Contents
A multi-agent system splits an industrial workflow among several AI agents, each with its own instructions, tools and permissions, coordinated by a manager agent or by code. It helps when the work divides into independent parts, such as gathering evidence from several plant systems. For strictly sequential tasks, a single agent or a fixed workflow is usually more reliable.
Key takeaways
- OpenAI and Anthropic both recommend getting the most from a single agent first, and adding agents only when the task requires it (OpenAI, Anthropic).
- In a controlled study, centralized coordination improved performance by 80.9% on a parallelizable financial task, while every multi-agent variant degraded performance by 39 to 70% on a sequential planning task (Google Research, 2026).
- Multi-agent systems are expensive to run: Anthropic reports they use about 15 times more tokens than chat interactions (Anthropic, 2025). On premise, that is GPU capacity.
- Failures are often about design and coordination: an analysis of more than 1,600 traces found 14 failure modes in 3 categories, system design, misalignment between agents and task verification (Cemri et al., 2025).
- In a plant, the strongest reason to split agents is to separate permissions: read-only specialists, one drafting agent, and approvals and control decisions kept with people.
What is a multi-agent system in an industrial setting?
Agents are not new to automation. The German VDI/VDE 2653 guideline series covers multi-agent systems in industrial automation, and its Part 4 (2022) gives "selected architectures for multi-agent systems (MAS) and selected patterns for field level control" (VDI). IEEE 2660.1 (2021) sets out recommended practices for connecting industrial agents to low-level automation functions (IEEE SA). Both concern software agents in automation and control, and both were published before agents built on language models became common.
This article is about the newer kind: agents built on large language models (LLMs). Each agent is a model with instructions and a set of tools that it chooses to call; the difference between a chatbot, a copilot and an agent is covered in chatbot vs copilot vs AI agent. A multi-agent system puts several of them together. Anthropic's research feature, for example, uses a lead agent that plans the work and creates subagents that search in parallel, then combines their findings (Anthropic, 2025).
Researchers are testing the idea on industrial tasks too. Agents4PLC uses LLM-based agents to generate PLC code and verify it at code level (Liu et al., 2024). For the broader use cases of agents in plants, see agentic AI for industrial process automation.
How do the agents coordinate?
Vendor documentation describes a small number of patterns. OpenAI separates a manager pattern, where a central agent calls specialists "via tool calls", from a decentralized pattern, where agents "hand off" the task to one another (OpenAI, A practical guide to building agents). Its SDK also allows orchestration in code, which makes speed, cost and performance more predictable (OpenAI Agents SDK docs). Anthropic describes the same ideas as workflows: orchestrator-workers, parallelization and evaluator-optimizer loops (Anthropic, Building effective agents).
| Pattern | How it works | Plant example | Watch out for |
|---|---|---|---|
| Fixed workflow in code | Code calls agents in a set order | Shift log → summary → formatted handover report | Rigid, but the easiest to test and audit |
| Manager (orchestrator-workers) | A manager agent calls specialists as tools and combines their results | Deviation investigation: one agent reads historian exports, one LIMS results, one maintenance history, one procedures | The manager is a bottleneck and a single point of failure |
| Handoffs (decentralized) | A triage agent passes the whole task to a specialist, which takes over | A plant helpdesk that routes a question to a maintenance or a quality agent | Context can be lost at each transfer |
| Evaluator loop | One agent drafts, another checks against criteria until they pass | A draft report checked against the site template and specification limits | Needs clear criteria and a maximum number of loops |
| Independent parallel agents | Agents work separately; results are merged at the end | Screening a large document set, one batch per agent | Errors are not cross-checked before merging |
Shift handover is covered in more detail in automating shift reports and handover.
When does splitting the work into agents help?
A controlled comparison comes from a Google Research study of single-agent and four multi-agent designs across 180 configurations (Google Research, 2026):
- On Finance-Agent, a task that splits into parallel parts, centralized coordination improved performance by 80.9% over a single agent.
- On PlanCraft, a strictly sequential planning task, every multi-agent variant degraded performance by 39 to 70%.
- Independent agents amplified errors up to 17.2 times; a centralized design contained this to 4.4 times, with the orchestrator acting as a check.
- Coordination overhead grew with the number of tools, so tool-heavy tasks need care.
Anthropic reports a similar pattern from practice. Its multi-agent research system outperformed a single agent by 90.2% on an internal research evaluation, yet the same article says domains that need shared context or have many dependencies between agents are "not a good fit for multi-agent systems today" (Anthropic, 2025). Cognition makes the opposite case for long tasks: "actions carry implicit decisions, and conflicting decisions carry bad results", so agents working in parallel on different assumptions produce work that does not fit together (Cognition, 2025).
The vendors' advice is consistent. OpenAI's general recommendation is "to maximize a single agent's capabilities first", and to split when instructions become too complex or tools overlap; some systems handle more than 15 distinct tools while others struggle with fewer than 10 overlapping ones (OpenAI). The LangChain documentation notes that "not every complex task requires this approach" (LangChain docs).
Which industrial workflows suit a multi-agent design?
Map the workflow before choosing. Workflows with independent branches suit several agents; chains where each step needs the previous result usually do not.
- Deviation and incident investigation. Specialists gather evidence in parallel from historian exports, lab results, maintenance history and procedures; a manager compiles a draft for the engineer. Each specialist is read-only.
- Document review across large sets. Checking a vendor data book or a contractor's documents against a specification, one agent per section, with a reviewing agent that checks the findings. See RAG over maintenance manuals.
- Periodic reporting from several systems. Monthly production, quality and maintenance summaries, where each agent queries one source. See text-to-SQL on production databases.
- Maintenance preparation. One agent reads asset history, another searches procedures, a third drafts a notification for a planner, as described in connecting an AI agent to a CMMS.
Poor candidates are sequential procedures such as start-up or shutdown steps, anything inside a control loop, and short tasks a single agent already does well.
What does a multi-agent system cost to run on premise?
Anthropic reports that agents use about 4 times more tokens than chat interactions, and multi-agent systems about 15 times more. In its analysis, token usage alone explained 80% of the variance in performance on the BrowseComp benchmark (Anthropic, 2025). Part of the gain from more agents comes from spending more compute.
On your own servers, tokens become GPU time and concurrent requests. A design with several agents can multiply the load of a single assistant, so size the hardware for the workflow, not for a chat demo; see on-premise LLM hardware and GPU sizing. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls (Staffing Industry Analysts, 2025).
Why do multi-agent systems fail?
The MAST study analysed more than 1,600 annotated traces from 7 multi-agent frameworks and found 14 failure modes in 3 categories: system design issues, misalignment between agents, and task verification. Its authors note that performance gains on popular benchmarks "are often minimal" (Cemri et al., 2025).
Anthropic's engineers list the practical problems: agents are stateful and "errors compound", runs are non-deterministic even with identical prompts, and a lead agent waiting for subagents creates bottlenecks. They added tracing, retries and checkpoints that resume from the point of failure (Anthropic, 2025). In a plant, the same problems show up as a wrong tag passed from one agent to the next, two agents reading different time windows, or a summary that hides a missing data source.
How do you keep a multi-agent system safe in a plant?
- One role, one tool list. Give each agent only the tools its role needs. Specialists that read data never hold write tools; only one agent drafts, and it creates drafts only.
- Authorization in the target system. The Model Context Protocol specification says tool servers "MUST" validate inputs, implement access controls, rate limit calls and sanitize outputs (MCP specification). The historian, LIMS or CMMS decides what each agent's account may do, not the model.
- Treat messages between agents as untrusted data. An agent that reads a supplier document can pass a hidden instruction to the next agent. Validate structured outputs between agents instead of passing free text.
- Human approval on actions. The MCP specification says there "SHOULD always be a human in the loop with the ability to deny tool invocations", and OpenAI recommends human oversight for sensitive, irreversible or high-stakes actions. Where to place these gates is covered in human-in-the-loop for AI agents.
- Stay out of control. CISA and partner agencies advise limiting active control of OT by AI without a human in the loop, preferring push-based or brokered architectures that move data out of OT, and failsafes that revert to traditional automation or manual operation (CISA et al., 2025). See the Purdue model and where AI fits and data diodes and one-way transfer.
- A distinct identity per agent. CISA recommends that the logged AI identity be distinct from other machine and user identifiers. With several agents, log each one separately so a wrong record can be traced to the agent that wrote it.
- Limits and tests. Set a maximum number of iterations, as Anthropic recommends (Anthropic, Building effective agents), and test on copies of real systems; see how to evaluate an LLM on your own documents.
Two open protocols are emerging for connecting the pieces. MCP standardizes how an agent reaches tools and data. A2A, developed by Google and donated to the Linux Foundation, standardizes how separate agents exchange tasks, and lets them work together "without needing to share internal memory, tools, or proprietary logic" (A2A protocol). Neither replaces the controls above.
What we have seen at KABAS AI
DALILE, our on-premise assistant, is built as a multi-agent system: a coordinator understands the question and routes it to specialized agents for search, data analysis and report generation. It runs on the client's own servers, connects to systems such as ERP, LIMS, document management and SQL databases, and includes access controls and audit logging. Our agentic AI team builds custom agents and workflows on it.
Our two published projects, at ADFC and ENAD Shymeca, use Virtual Lab for R&D; both companies are exploring integration with DALILE. So this is the sequence we recommend when scoping a multi-agent workflow, not a reported result:
- Draw the workflow and mark which steps are independent and which depend on the previous one.
- Build it with one agent first and measure it on real cases.
- Split only where the evidence points: parallel evidence gathering, too many overlapping tools, or a need for separate permissions.
- Agree each agent's tool list and approval points with operations, maintenance and HSE before writing code.
Sources
- Anthropic, How we built our multi-agent research system (June 2025)
- Anthropic, Building effective agents (December 2024)
- OpenAI, A practical guide to building agents
- OpenAI Agents SDK documentation: Agent orchestration
- LangChain documentation: Multi-agent
- Google Research, Towards a science of scaling agent systems: When and why agent systems work (January 2026)
- Cemri et al., Why Do Multi-Agent LLM Systems Fail?, arXiv 2503.13657, 2025
- Cognition, Don't Build Multi-Agents (June 2025)
- Liu et al., Agents4PLC: Automating Closed-loop PLC Code Generation and Verification in Industrial Control Systems using LLM-based Agents, arXiv 2410.14209, 2024
- VDI, VDI/VDE 2653 Part 4: Multi-agent systems in industrial automation, selected patterns for field level control and energy systems (2022)
- IEEE SA, IEEE 2660.1-2020, Recommended Practice for Industrial Agents: Integration of Software Agents and Low-Level Automation Functions
- CISA, ASD's ACSC and partners, Principles for the Secure Integration of Artificial Intelligence in Operational Technology (December 2025)
- Model Context Protocol specification (2025-11-25), Tools
- A2A Protocol documentation
- Staffing Industry Analysts, Gartner says agent washing is taking place (June 2025)
Frequently asked questions
What is the difference between an AI agent and a multi-agent system?
An AI agent is one language model that plans its next steps and calls tools to reach a goal. A multi-agent system splits the work among several agents, each with its own instructions, tools and context, coordinated by a manager agent, by handoffs between agents, or by fixed code. The individual agents work the same way; the difference is how the work and the permissions are divided.
Is a multi-agent system always better than a single agent?
No. In a controlled study by Google Research, centralized coordination improved results on a parallelizable financial task, but every multi-agent variant tested lowered performance on a sequential planning task. OpenAI and Anthropic both advise getting the most out of a single agent first and adding agents only when the task requires it.
Can a multi-agent system control a plant process?
It should not act directly on control systems. Joint guidance from CISA and partner agencies advises limiting active control of OT by AI without a human in the loop and pushing data out of the OT network rather than granting AI systems inbound access. Agents can read exported data and prepare recommendations that go through the site's existing change management.
Should agents talk to each other directly or through an orchestrator?
For most industrial workflows, through an orchestrator. A central manager agent keeps one view of the task, checks the specialists' results and is the single point where approvals and logs are enforced. In the Google Research study, independent agents that did not communicate amplified errors far more than a centralized design.
Do we need MCP or A2A to build a multi-agent system?
No. The Model Context Protocol standardizes how an agent connects to tools and data, and A2A standardizes how separate agents exchange tasks and results. They help when agents and tools come from different teams or vendors. Inside a single on-premise application, agents can also be coordinated in code, and the controls are the same: narrow tools, access checks in the target system, logs and human approval.
How do we test a multi-agent system before using it in operations?
Run it on realistic scenarios against test copies of the systems it reads and writes, several times per scenario, because agent runs are not deterministic. Trace every agent's tool calls and messages, judge the final records rather than the agents' messages, and compare the result with a single-agent version of the same workflow.



