Dynamic process models are increasingly consumed by automated agents rather than by control engineers reading source code. These agents include data pipelines that assemble training sets, reinforcement-learning algorithms, and language models. Each depends on process meaning that a numerical interface does not expose, such as what each variable represents, which signals are available online, and which limits trigger a shutdown. Today that meaning lives in source code and informal convention, so recovering it is error-prone and does not transfer across a model’s many independent implementations.

The design principle is that a simulator should be self-describing. It publishes one machine-readable process description, and that description is the single source from which every interface is derived.

A single machine-readable process description sits above the process-model core and below four interfaces for online interaction, trajectory data, estimation and optimization, and agent access, each serving people, algorithms, or reasoning models.

One description, many views. The transition core advances the process. The online-interaction, trajectory-data, estimation and optimization, and agent (MCP) interfaces are all derived from the same description, so engineers, algorithms, and reasoning models act on one definition of the process.

Reference implementation

TEP Studio applies the principle to a modified Tennessee Eastman process, a standard testbed for three decades because it captures a real plant’s nonlinearity, coupling, and disturbance structure.

  • Fast. A self-contained Python implementation simulates one process-hour in about 8 ms. That is more than five times faster than the Fortran reference and about two orders of magnitude faster than an existing Python implementation.
  • RL-ready. It is a Gymnasium environment, so reinforcement-learning and control methods can train on it directly.
  • Agent-accessible. The process description is exposed through a Model Context Protocol (MCP) server, giving language models structured access to the same semantics.

The interfaces for online control, trajectory data, estimation and optimization, and agent access are all generated from the same description, so a single definition is shared instead of restated for each consumer.

The code is at github.com/khalidlabs/tep-studio, and a live simulator runs at khalidlabs.com/tep.