Agentic workflow with verification for tribology simulations

Artificial Intelligence is taking the world by storm at unprecedented speed. The associated changes are going to impact the way of working for most of the people, including tribologists and simulations engineers. Although the future is only being shaped right now, currently, it is believed that the combination of humans and AI agents will deliver the most gains in productivity and overall wealth creation. In line with this narrative, I have been working on the use of OpenClaw (and Hermes now) to support my tribology simulation work with TriboSolver. Few use cases were considered and the summary is as follows:

• AI agents are capable of setting up, running and reporting complex simulations is such tools like TriboSolver (any other tool with web interface or API would also work). They can also help in preparation by generating inputs (e.g., surface roughness)
• Verification step is necessary to avoid hallucination results. This step could be as simple as checking the existence of the results file
• Humans are still needed to define the direction and verify the final result, at least at the current stage of affairs
• AI agents can be guided via interfaces that are used by humans, e.g., Telegram chat (including voice instructions). The user is not required to seat in front of the computer to perform such work. Consequently, in future (not so far away) we will not use UIs to work with such tools and computers in general, we will use AI agents instead and interact with them like with humans

The verification bullet point is the key topic of the article. In the previous simulation, it was found that the AI agent confidently hallucinated pressure map as something went wrong with the retrieval of the information from the results file. That is a known problem for AI tools, and it has to be kept in mind when working together with AI agents as well. To address this issue and reduce the amount of rework needed from a human, it is proposed to use a second, independent AI agent to verify that the main simulation agent did exactly what it was told to do. This additional step is intended to check meta data of the simulation, e.g., the existence of the simulation result data file, the time of the simulation, the inputs for the simulation, as well as quick check of the results. This step is verifying that the outputs correspond to outputs and nothing more. This step would eliminate the issue encountered in the previous case, when the hallucinated pressure map was delivered as the calculation result. At the same time, it must be emphasized, this check will not tell, whether the simulation is physically correct or not. This part remains for the human to verify. The second AI agent verification is to exclude some simple mistakes in the analysis, not to completely verify the soltuion.

The verification workflow is shown in the figure below. As the second AI agent, I used Hermes, but it could also be any other independent AI agent, including OpenClaw subagent. 

The problem set up was the same as in the previous case, where contact stiffness of a sinusoidal and random rough surface was calculated as a function of load. The difference in this case is that the problem set up was performed via a Telegram chat which included both AI agents, OpenClaw and Hermes. OpenClaw agent was performing the work, while Hermes agent was veifying the results.

What was simulated

The case used TriboSolver dry contact simulation with the Boundary Element Method. Two rough surfaces were compared across five normal loads: 0.02, 0.05, 0.10, 0.20 and 0.40 N.

• Sinusoidal surface: 50 nm amplitude and 4 µm period.
• Random fractal rough surface: Sq = 10 nm, H = 0.8.
• Extracted quantities included calculated load, indentation, real contact area, average contact pressure, tangent stiffness and PADIS pressure maps.

Pressure maps

The pressure maps below were extracted from TriboSolver result files through the Results callbacks. They show the lowest and highest load levels for both surfaces.

Why a second verification agent is needed

Agentic simulation is useful only if the result can be trusted. A browser-controlling agent can click the wrong option, reuse an old result file, extract a metric from the wrong dropdown, or report a plot that was generated from a failed callback. None of these errors look dramatic in a chat transcript. They usually look like a normal answer.

That is why this workflow separates execution from review. OpenClaw runs the simulation and prepares the evidence package. Hermes checks the package independently. The human still makes the final engineering judgement.

In this case Hermes checked that all 10 cases were present, the calculated loads matched the requested loads, indentation and real contact area increased monotonically with load, stiffness stayed positive, real area stayed below nominal area, and the pressure maps came from the expected TriboSolver outputs.

What the results showed

Both surfaces stiffened with increasing load. The tangent stiffness rose from roughly 1.8-2.0 MN/m at 0.02 N to roughly 4.2-4.3 MN/m at 0.40 N.

The random fractal surface produced a larger real contact area at all loads and therefore much lower average contact pressure. The sinusoidal surface was more localized, with average pressure rising from about 499 MPa to about 1118 MPa over the same load range.

Download the report

The full PDF report, including the Hermes verification table and all extracted result plots, is available here: TriboSolver final report with Hermes verification.

Why this matters

The goal of introducing a second verification agent is to remove repetitive friction from the workflow while adding more checks than a rushed manual process usually gets. TriboSolver remains the solver and the source of truth. OpenClaw automates the browser workflow and report assembly. Hermes reviews the evidence. The final decision stays with a person who understands contact mechanics.

For research, consulting and technical education, this is a useful pattern: run the calculation through the actual engineering tool, preserve the artifacts, verify the extraction, and publish only after human review.