##  [06. PyRocky as an Execution Layer for Agentic AI](/blog/06-pyrocky-execution-layer-agentic-ai) 

**Agentic AI** represents a further evolution of **AI-assisted**engineering.

While a conventional **AI** model generally receives an input and produces a prediction, an **AI** agent can be designed to plan and execute a sequence of actions toward an engineering objective.

In future or user-developed agentic workflow, **PyRocky** could serve as the programmable execution layer between an **AI** agent and **Ansys Rocky**.

For example, an engineer could define an objective such as:

- *Evaluate the transfer chute performance between 800 and 1,200 tons per hour and identify operating conditions that maintain stable discharge while minimizing maximum particle impact velocity.*

An engineering **AI agent** could then be configured to:

- Interpret the requested engineering objective.
- Select the **Ansys Rocky** parameters that should be varied.
- Generate the required simulation cases.
- Use **PyRocky** to launch and control the simulations.
- Monitor simulation execution and identify failed cases.
- Extract mass flow, velocity, collision, and wear-related results.
- Compare the results against predefined engineering constraints.
- Select the next simulation points using an optimization strategy.
- Produce a technical summary of the evaluated configurations.
- Present recommendations for review by the responsible engineer.

In this architecture, the **AI** agent would coordinate the workflow, while **PyRocky** would provide controlled access to **Ansys Rocky’s**simulation capabilities.

This combination could reduce the time required to investigate large design spaces and enable more interactive engineering workflows based on natural-language objectives.

However, such systems must be developed with appropriate engineering controls.

**AI-generated** actions should be restricted by validated parameter ranges, predefined unit systems, approved simulation templates, and explicit acceptance criteria. All automatically generated models should preserve a complete record of input parameters, software versions, solver settings, warnings, and results.

Most importantly, engineering validation must remain a human responsibility. An **AI agent** may assist with case generation, execution, data analysis, and recommendation, but it cannot independently guarantee that:

- The selected contact model is physically appropriate.
- Material parameters have been properly calibrated.
- Particle shapes and size distributions represent the real material.
- Boundary conditions accurately reproduce the equipment's actual operation.
- The numerical time step and solver settings are adequate.
- The simulation has reached the required physical regime.
- The resulting design is safe or suitable for industrial implementation.

For this reason, the most reliable architecture is a human-in-the-loop workflow in which the **AI agent** assists the simulation engineer, while the engineer remains responsible for model validation, physical interpretation, and final design approval.

---

- [Previous: 05. Foundation for AI-Enabled DEM Workflows](https://developer.synopsys.com/blog/05-foundation-ai-enabled-dem-workflows)