Hi folks,
Previously, an evaluation experiment in Agent Mesh pointed at a single agent: send a prompt, score the answer. Workflows are a different shape. They take structured input that has to match an input schema, and they run a graph of agents, tools and decision nodes. Now an experiment can target a workflow directly.
A few things make that work. Dataset examples can carry input data as a JSON object, next to a prompt or instead of one. The target is checked when you save the experiment, not halfway through a run, so a misspelled workflow name or an example missing a required field is refused up front, and the error tells you which example and which property. You can also add up to five evaluators to any experiment, and max workers is enforced so a big run doesn’t swamp the mesh.
Each trial reports the workflow’s output, its scores, and token usage broken down by node and by model. It also tells you why the workflow ended and which node failed, and each node’s result carries a short preview of what it saw. That last bit saved me a lot of guessing when I was debugging a decision node.
In the video I build the experiment with sam config apply, run it through the platform API, and read the results in the web UI, including a real failure:
I’m curious how you’re testing your workflows today. Which workflow would you point an experiment at first? Drop it in the comments.
You can download Solace Agent Mesh and find the repo on GitHub: