Short answer
In industrial companies, coding agents deliver the fastest gains in IT and integration code, solid gains in test and documentation work on legacy systems, and need strict boundaries near OT and safety-relevant embedded software. A workshop for manufacturers should reflect those zones instead of treating all software the same.
On this page
Industrial software is not one thing
A typical manufacturer runs several software worlds at once. There are ERP and MES integrations, data pipelines from the shop floor, internal web tools, PLC programs, embedded firmware in products, test benches, and decades-old applications that still run a critical line. The people writing that software range from enterprise developers to automation engineers who learned to code on the job.
Coding agents behave very differently across these worlds. They are productive in a TypeScript dashboard, helpful in a C++ test harness, and potentially dangerous if someone lets them modify safety-related control logic without the usual verification. A useful workshop starts by drawing those lines. That is how we approach industrial sessions at neooutput.
The three zones model
We use a simple model that participants grasp quickly and that safety and quality managers accept:
| Zone | Examples | Agent use in the workshop |
|---|---|---|
| Green: IT and integration | ERP connectors, internal web apps, data pipelines, reporting | Full agentic workflow with normal review |
| Amber: engineering support | Test benches, simulation code, tooling, documentation, legacy analysis | Agents for analysis, tests and docs; changes reviewed by domain experts |
| Red: control and safety | PLC logic in production, safety functions, certified firmware | Read-only analysis, documentation and test ideas only; no autonomous changes |
The zones are a starting point. Each company adjusts them to its own standards, certifications and risk appetite. Writing them down during the workshop is already a valuable outcome.
What agents do well in industrial settings
Understanding legacy code nobody owns
Many plants depend on applications written by people who have since retired. Agents are good at reading such code, explaining it, drawing call graphs in text form and writing characterization tests that capture current behaviour. That reduces the risk of every later change. For larger modernization projects, our legacy modernization workshop covers the full approach.
Integration glue
Moving data between machines, MES, ERP and analytics platforms involves a lot of mapping code, format conversions and error handling. Agents handle that well, especially when the team provides example messages and a clear target schema.
Test and simulation code
Test harnesses, mock devices and simulation scripts are ideal agent tasks: well specified, verifiable and outside the production control path.
Documentation for audits and handovers
Quality management systems require documentation. Agents can draft interface descriptions, change logs and module documentation from code, which engineers then check and approve.
Where to be careful
- Vendor-specific PLC languages and toolchains. Model knowledge of proprietary environments varies widely. Expect more errors and verify everything.
- Hardware in the loop. An agent cannot see the machine. Never let it run commands that move actuators or change live parameters.
- Certified software. Where functional safety standards or customer certifications apply, AI-assisted changes must go through the full verification process. The workshop should not suggest shortcuts.
- Network segmentation. Development machines that can reach OT networks need clear rules about what an agent may execute.
A two-day format for mixed teams
Day one: shared foundations
The morning brings IT developers and automation engineers together. We introduce coding agents, the three zones and practical safety rules. Then participants work on green-zone tasks in mixed pairs, which helps automation engineers learn agentic workflows on lower-risk code.
In the afternoon, pairs move to amber-zone tasks: analysing a legacy module, generating tests for a test bench or documenting a communication protocol from code. Each pair writes a context file for its repository, including explicit statements of what the agent must not touch.
Day two: tracks by role
Track A for IT and integration teams focuses on multi-file features, CI integration and review habits.
Track B for automation and embedded engineers focuses on analysis, test generation in C or C++, documentation and working with toolchains that run outside the agent's reach. We show how to use the agent as an assistant that proposes changes for the engineer to apply and verify in the vendor environment.
The day ends with a joint session on governance: zones, approved tools, data rules and who owns updates.
Data and IP considerations
Industrial source code often contains valuable know-how about processes and products. Before the workshop, clarify with management and IT security:
- Which repositories may be used with which tool and under which contract terms
- Whether any code must stay entirely on premises, which may point to self-hosted options
- How to handle customer-specific code under non-disclosure agreements
Beyond engineering
Once engineering teams use agents successfully, other parts of the company usually want AI agents too: quality reports, maintenance documentation, supplier communication. Those use cases need governance at a different level, with audit trails, SSO and integrations into SAP and other systems. kalAIDoCreate provides that as an orchestration platform with an expert team. Our article on connecting AI agents to SAP safely describes typical integration patterns.
neooutput runs industrial workshops on site at plants or engineering centres as well as remotely, in German or English, and tailors exercises to your actual stack. For leadership teams deciding how to organize all of this, the executive workshop for CTOs and tech leads is a good starting point.
FAQ
Can coding agents write PLC code?
They can produce code in some PLC languages, but quality varies and verification in the vendor toolchain is essential. In production and safety contexts we recommend using agents for analysis, documentation and test ideas rather than autonomous changes.
Is it safe to use cloud-based agents with our source code?
That depends on your contracts, data classification and the provider's terms. Many manufacturers start with integration and tooling code, then decide on broader use after a review.
Do automation engineers without classic software training benefit?
Yes. The agent's ability to explain code and propose tests is particularly helpful for engineers who code as part of a broader role, as long as review and zones are clear.
How do we prevent agents from affecting live systems?
Keep development environments separated from OT networks where possible, restrict which commands agents may run and never allow direct access to production controllers.