← Back to digest
AI Agents

SemaPLC: verification-gated code for industrial controllers

What's this about?

Programmable logic controllers run industrial plants, and language models can already generate individual program organization units for them. This paper's question is the one that matters in practice: whether that generated logic integrates into an existing PLC project without breaking it. SemaPLC is a project-grounded, verification-gated agent harness — generated code must pass checks against the real project before it lands.

What exactly is a PLC, and why is the bar different?

A programmable logic controller is the industrial computer that runs physical processes — a bottling line, a water treatment plant, a furnace. Its code operates on a strict cycle, reading sensors and setting outputs continuously. The consequences of a bug are not a stack trace: a valve opens when it shouldn't, a press cycles with someone's hand in it. So the industry has decades of certification practice, and "the model wrote something that compiles" is nowhere near the bar. Verification-gated means the harness refuses to accept code that hasn't passed those checks — the enforcement lives outside the model, which is the same principle as Mandato and PACE in other domains.

Why is this worth noticing?

111 upvotes
  1. It's agentic coding entering a domain with real safety consequences, where the informal review practices of software development are not acceptable.
  2. The project-grounded framing is the useful contribution. Benchmarks generate isolated units; the actual difficulty is integration into a large existing system with its own conventions and interlocks.
  3. It's the same architecture as this week's other safety work — capability from the model, enforcement from the harness — arriving now in industrial control.
← Zetta: a closed-loop embodied harnessCo-RL: reasoning without ground truth →