Answer Hub
AI Agents for Manufacturing: Everything You Need to Know
How AI agents find the next machine failure in the data your plant already has, and how to run them inside the plant without giving up control, in one place.
Chapter I
Introduction
Most plants already collect the data that shows the next machine failure coming. It sits in the MES and the historian, and the people who can read it are retiring or stretched thin. AI agents can read that data on every shift. The Giant Swarm Agent Platform runs those agents inside the plant's own environment, under rules IT sets once, so they help without taking control away from IT and OT.
AI Agents for Manufacturing Key Takeaways
AI agents in manufacturing read the data a plant already collects, such as downtime reports, sensor trends and production schedules, and turn it into proposals like a repair window before a machine fails. Run safely, they stay inside the plant, act only with the permissions of the person using them, and change nothing in a production system until a person approves.
Chapter II
Who is this for?
This page is for the people who decide whether AI agents get access to plant systems, and who answer for it afterward. It's especially relevant if you are a:
- Chief Information Officer (CIO) or Chief Digital Officer (CDO) deciding how AI fits into the plant's architecture, data and risk posture
- Head of IT/OT responsible for connecting production systems without opening new security gaps
- Chief Operating Officer (COO) looking to reduce unplanned downtime without adding operational risk
- Director of Manufacturing or Maintenance who needs repairs planned around production, not after a breakdown
If your job involves plant data, uptime or the systems that run the line, you're in the right place.
Chapter III
Why it matters
Unplanned stops are expensive, and most plants already have the data to see them coming. Downtime reasons sit in the MES, sensor trends in the historian, and the production schedule in the planning tools. What's missing is the time to read them on every shift: the technicians who know how to diagnose a fault are retiring, and IT and OT teams are short of people. In Germany, 78% of companies report a shortage of IT specialists, according to Bitkom's September 2026 labour market survey.
AI agents can close that gap, and companies are starting to use them: 11% of German companies already use AI agents and another 29% plan to, according to Bitkom's September 2026 survey of 603 companies. In a plant, though, an agent touches production systems, so the questions change. Plant data often must not leave the site, which is one of the main reasons to process data close to where it is produced, at the edge. Sites lose their network connection. And nothing may change in a production system without someone accountable for it.
Security agencies now publish guidance specifically for AI in operational technology (OT). In December 2025, CISA and Australia's Cyber Security Centre, together with partners including Germany's BSI, published joint guidance on the secure integration of AI in OT. It sits next to the established OT security references: NIST SP 800-82, the Guide to Operational Technology Security, and the ISA/IEC 62443 series for industrial automation and control systems.
Treating the plant as a special case makes all of this harder. When tools treat the edge as an exception, pipelines fork, monitoring breaks and policies drift; treating the edge like infrastructure is what keeps many sites manageable.
Chapter IV
How Giant Swarm approaches AI agents in manufacturing
Giant Swarm has run production platforms for manufacturers for more than twelve years. The Agent Platform applies the same operational discipline to AI agents in the plant. The approach rests on a few ideas:
- The platform runs in your environment. On your hardware or in your cloud, with an open-weight model if you want. With a model on your own hardware, no plant data leaves the site, and sites without an internet connection get releases through an offline channel.
- Every write waits for a person. Agents read from the MES, the historian and the shift plan. Before they write to the ERP or the maintenance system, they show the exact call and wait until a person allows it. The approval and the call land in the audit trail.
- Agents act with the permissions of the person who started them. An agent never gets more access than the engineer who uses it, and IT approves every connector once. Connectors use the open Model Context Protocol.
- Built once, reused in every plant. A forward deployed engineer sets up the platform and builds the first agents with your team on one line. The next line reuses the agents and connectors, and the second plant starts from what the first one built.
- Cost is visible per line, team and model. Each team has a budget, and IT sees every model call.
- Yours to keep. The platform is open source, Apache 2.0, with components in the CNCF and the Linux Foundation. It runs on the Kubernetes you already have, including Kubernetes on bare metal, and stays with you if the contract ends.
Chapter V
FAQs
How do factories safely introduce AI agents into existing production systems?
Start read-only, and keep a person in charge of every change. On the Giant Swarm Agent Platform for manufacturing, agents read from the MES, the historian and the shift plan, and they act with the permissions of the person who started them, never more. When an agent wants to write to the ERP or the maintenance system, it shows the exact call and waits until someone allows it, and the approval lands in the audit trail. IT approves each connector once, so plant teams build agents from building blocks that are already cleared. That way the first agents create value on a single line without anyone handing production control to a model.
Can AI agents run without plant data leaving the site?
Yes. The Agent Platform runs in the plant's own environment, on its own hardware or in its own cloud, and with an open-weight model on that hardware no plant data leaves the site. That matters because privacy rules and regulations often require that certain data never leaves the premises, one of the core reasons edge computing exists in the first place. Sites without an internet connection get platform releases through an offline channel, so the agents also work in plants that are fully cut off from the outside.
Which plant systems can the Agent Platform connect to?
The systems a plant already runs. Agents read orders, line status and downtime reasons from the MES, sensor trends such as bearing temperature from the historian, and qualifications from the shift plan. After approval, they write spare parts reservations to the ERP, for example SAP PM, and work orders to the maintenance system. Connectors use the open Model Context Protocol, so teams reuse existing ones, and a forward deployed engineer builds the missing ones with the plant team.
How do AI agents help with predictive maintenance?
They do the reading nobody has time for. In the manufacturing demo, which uses fictional data, a Production Analyst finds the machine that lost the most time and the cause, here a worn drive bearing. A Maintenance Planner checks the production schedule, the spare parts store and the crew, and proposes a repair window that costs no customer order. A Work Order Assistant prepares the work order and waits for one approval before anything is booked. The know-how of experienced technicians can go into the agents' instructions, so it stays in the plant when people retire.
How do manufacturers scale AI agents from one line to every plant?
By building once and reusing. A Giant Swarm forward deployed engineer sets up the Agent Platform in your environment and builds the first agents with your team on one line. The next line reuses those agents and connectors, and the planner adapts the instructions to its machines. The second plant starts from what the first one built, without a second setup. The first line starts with the Foundation Package at a fixed price.
Which security standards apply to AI agents in OT environments?
The established OT security references still apply: NIST SP 800-82, the Guide to Operational Technology Security, and the ISA/IEC 62443 series, which defines requirements and processes for secure industrial automation and control systems. For AI specifically, CISA and Australia's Cyber Security Centre, with partners including Germany's BSI, published joint guidance on the secure integration of AI in OT in December 2025. A platform that logs every agent action and ties it to a person makes it easier to show who did what in a production system.
What happens when a plant loses its network connection?
The plant keeps running. Vanilla Kubernetes treats a node that drops off the network as failed and reschedules its workloads after five minutes, which is the wrong behavior at the edge. Projects like KubeEdge let edge nodes keep running their workloads on their own and sync back once the connection returns. Giant Swarm designs for disconnected sites: configuration is pulled by the edge rather than pushed, and the Agent Platform gets its releases through an offline channel where there is no internet connection.
How do you keep the cost of AI agents under control across plants?
Make it visible from the start. On the Agent Platform, cost is visible per line, per team and per model, each team has a budget, and IT sees every model call. Because agents and connectors are built once and reused, the second line and the second plant don't pay for a second setup. And with open-weight models on the plant's own hardware, inference doesn't depend on a hyperscaler's per-token pricing.
Do we need Kubernetes in the plant before we start?
No. The Agent Platform runs on the Kubernetes you already have, wherever it runs. If you don't run Kubernetes yet, Giant Swarm provides it, including Kubernetes directly on bare metal without a hypervisor layer. For edge sites, Giant Swarm treats the edge as part of the platform, so a factory site gets the same lifecycle, compliance and control as the data center.
Closing thoughts
AI agents can find the next failure in data a plant already has, but only if IT and OT can trust them. That means agents that run inside the plant, act with a person's permissions and wait for approval before they change anything. Giant Swarm brings more than twelve years of running production platforms for manufacturers to that job.
Talk to our team or see the Agent Platform for manufacturing to try it on one of your lines.


