Manufacturing

AI agents for robot programming read the program, not the cell

AI agents for robot programming can only improve the code they can see. HD Hyundai's 63% is real progress, but the cell's history lives elsewhere.

8 min read

HD Hyundai Robotics said on 1 November that its CTO had presented results at Palantir’s Sovereignty Bootcamp in New York: average task time down from 21.5 days to 7.9 since AI agents joined the work in August. It’s the most concrete number I’ve seen for AI agents for robot programming. The agents read existing robot programs and suggest improvements. The report doesn’t ask the obvious question. How much of what a robot actually does is in the program?

TL;DR: AI agents for robot programming are software that read existing robot or controller code, propose changes and, in the newer tools, carry out engineering tasks for an engineer to review. HD Hyundai reports a 63% drop in average task time, and Siemens and Rockwell are shipping or previewing agents for the PLC side. All of them work from the program and the project file. A running cell also depends on mastering data, frames, I/O and years of touch-ups that often sit outside any record. Get the stored program to match the live one before you measure what an agent saves.

What HD Hyundai actually reported

The account comes from The Elec, in a piece by Park Jong Woo dated 2 November 2026. I couldn’t find a second outlet with the figures, so treat it as a company statement made at a partner’s event.

The work has three parts. AI agents analyze existing robot programs and flag possible improvements. On-site engineers refine those suggestions from experience. And those engineers are being retrained as field solution specialists, people who define the problem and then use AI to implement the fix.

Two numbers came with it. Average task completion fell 63%, from 21.5 days to 7.9, since the agents entered the problem-solving process in August. And the company expects to cut robot programming costs by more than 50%.

Those are different kinds of number. The first is a reported result over roughly three months. The second is a forecast.

The first also has a hole in it. The report doesn’t say which tasks were counted, how many there were, or whether the mix changed. And the agents arrived in the same partnership that redefined the engineers’ job, one senior enough to plan regular executive steering committees. A cut of nearly two-thirds in elapsed days could come from the agent, from the new role, or from a project that finally had an executive’s attention. The report can’t separate those, so nobody outside the company can either.

I’d bet on the role change doing more than the headline implies. Few of those 21.5 days are typing. They’re usually waiting for a window on the line, finding who knows the cell, and working out what the last person changed. I can’t prove that for HD Hyundai. It’s the question I’d ask first.

The program is not the robot

Here’s the part that matters if you’re buying this. An agent that reads robot programs is reading one input to the machine’s behavior.

Take a FANUC cell, because its failure modes are well documented. Patti Engineering’s backup guide, updated 30 September 2026, warns that many sites make partial backups without realizing it. They save the teach pendant programs and miss the system configuration: mastering data, I/O assignments, system variables. A controller restored that way can run the same program and position inaccurately.

Industrial Monitor Direct’s deployment notes make the dependency plain. Vision calibration depends on correct user frames. Frames depend on a validated tool center point. Restore SYSMAST.SV from another robot and you get severe positioning errors that need a full re-mastering. Same program. Different robot.

Then there’s the history. A fixture wears, a gripper gets swapped, a conveyor moves 4 mm during a shutdown. Someone on nights nudges a point on the pendant and the line runs again. That fix is now the truth, and the offline simulation model doesn’t know about it. Those same deployment notes tell you to document touched-up positions. Everyone agrees with that advice at 3 a.m. and then goes home.

The survey data says this isn’t a rare case. In Copia Automation’s 2024 State of Industrial DevOps survey of 200 US executives, 79% called ad hoc fixes in industrial programming commonplace. The release notes those changes often aren’t tracked. Respondents also reported spending 45 hours a month debugging code, about ten times what they spent reviewing it. Copia’s second annual report attributes 45% of downtime to industrial code issues. Copia sells code management, so read those as a vendor’s numbers. They still match what anyone who’s walked a line would guess.

Now put an agent on that. It reads a program, sees a redundant move or a wait that looks pointless, and proposes removing it. Maybe that wait exists because a clamp sensor on station 3 has been slow since 2022, and the person who added it has left. The code is cleaner. The cell is worse.

This is a neighbor of the problem I wrote about in August, when AI agents operating equipment needed a spec nobody wrote. That piece was about what a machine may be told to do. This one is about what’s already been done to it. A good agent makes that gap faster to hit, for the same reason a redesigned room is what lets a robot work at all: the structure around the machine carries more of the outcome than the demo admits.

Vendors are starting from the design side

ARC Advisory draws the line between the old tools and the new ones this way: a copilot suggests, while an agent works across tools and project data to finish a defined task under an engineer’s supervision. Rockwell’s launch FAQ makes a similar contrast. So look at what the big automation vendors connect those agents to first.

ToolWhat the agent works fromReported resultStatus of the evidence
HD Hyundai Robotics with PalantirExisting robot programs, refined by on-site engineersAverage task time 21.5 to 7.9 days; more than 50% lower programming cost expectedCompany statement, reported by The Elec, 2 Nov 2026
Siemens Eigen Engineering AgentThe live TIA Portal project: devices, PLC blocks, data types, HMI screensMass parameter change across 66 PLC blocks from 6 minutes to 1.5; analyzing and documenting 200 lines of SCL from 35 minutes to 11Siemens examples given to ARC Advisory, which calls them use-case evidence
Rockwell FactoryTalk ConductorStudio 5000 Logix Designer, Logix Echo, Emulate3D and FactoryTalk Optix at launchNone publishedAnnounced 21 Oct 2026; preview at Automation Fair 2026; commercial release expected Q1 2027

Siemens launched Eigen in April 2026 after piloting it with more than 100 companies in 19 countries, and its results are the most believable here because they’re small and specific. Bulk edits and explaining old code are exactly where this works. ARC’s view is that these tools are “not substitutes for engineering judgment”, and it lists the things that still need validating: motion sequences, safety-related behavior, commissioning realities.

Rockwell’s roadmap is the tell. Design, development, simulation and test tools come first. FactoryTalk Historian, Plex and Fiix are planned to follow. Historian holds what the process did. Fiix holds what maintenance did about it. The systems that know what happened to the machine after commissioning are the later integrations.

That order makes sense for a vendor. Project files are structured and they own them. It also means the first generation of these agents knows the design better than the plant does, and the plant is where the 3 a.m. fix lives.

The research sits in the same place. A May 2026 review by Vincent Henkel and colleagues screened 2,341 publications and kept 88 on foundation-model agents in industrial automation. It put 75.0% of reported systems at prototype and early validation stages, with deployment-oriented evidence at 9.1%. Agents4PLC, a research framework, checks generated PLC code at the code level against formal specifications. That’s useful, and it shows the code meets the spec. It can’t show the spec describes the fixture that’s bolted to your floor.

What to do before the agent

None of this argues against the tools. Explaining a 15-year-old program that nobody on staff wrote is a real gain. But the sequence matters, and the first steps are cheap.

Start by finding out what’s running. Pull a full backup from every controller, including mastering, I/O and system variables, and compare it with what your files say is live. Name the backups by robot ID, controller model, software version and date, as Patti Engineering suggests, so a restore doesn’t turn into archaeology. The diff is your first deliverable. It’s usually a surprise.

Then give changes a place to land. Put the programs under version control and require a one-line reason on every touch-up. It doesn’t need a platform on day one. It needs the second-shift lead to do it, which means it has to take less time than the fix.

Tie the program to the cell’s history. When maintenance swaps a gripper, the work order should point to the robot and the program. This is the same asset record problem one layer further in, and it’s the part an agent can’t reconstruct by reading code. Someone has to write down the work nobody documented.

Record a baseline before the pilot. Define when a task starts and when it ends, then time a few months of them. Without that clock you’ll get a number like 21.5 to 7.9 and no way to say what produced it. The same trap showed up in software, where coding agents sped up writing and the review queue absorbed the gain.

Last, keep the gate physical. Agent output goes to an engineer, then a reduced-speed teach-mode run, then production. Nothing writes to a live controller on its own.

This is the process and data-contract work that decides whether a programming agent shortens a job or just ships a confident change to a cell it never met.

HD Hyundai’s engineers refine what the agent suggests from experience. Right now, that experience is the only copy of the cell’s history.

FAQ

Can AI agents program industrial robots?
Partly. Today they review and change code for an engineer to check. HD Hyundai Robotics says its agents analyze existing robot programs and suggest improvements that on-site engineers refine. Siemens and Rockwell are building agents for automation engineering around PLC projects. A May 2026 review of 88 publications on industrial agents put 75.0% of reported systems at prototype or early validation stages and only 9.1% at deployment-oriented evidence. Unattended programming of a production cell isn't what these reports describe.
What is the difference between an AI copilot and an AI agent in automation engineering?
A copilot suggests. An agent works across tools and project data to finish a defined task under an engineer's supervision. That is ARC Advisory's distinction in its analysis of Siemens' Eigen Engineering Agent, and Rockwell's FAQ for FactoryTalk Conductor makes a similar contrast. ARC still keeps the engineer in control.
Did HD Hyundai's AI agents cut robot programming time by 63%?
The reported figure is narrower than that. The Elec reported on 2 November 2026 that average task completion time fell 63%, from 21.5 days to 7.9, since AI agents entered the problem-solving process in August. The report doesn't say which tasks were counted or how many. The expected cut of more than 50% in robot programming cost is a forecast, not a result.
How do you prepare robot programs for AI agents?
Make the stored program match the running one. Take full controller backups that include mastering data, I/O assignments and system variables, compare them with what the team believes is live, log every touch-up with a reason, and record a baseline for how long tasks take today. Then keep a person and a reduced-speed teach-mode check between the agent and the production cell.