Enterprise agents are complex systems, and they only pay off when their work is reliable and economically viable. Evaluation is how you get both: checking correctness is the obvious case, and routing is the subtler one, since choosing a model against cost, latency, and quality requires quality to be measurable at all. Knowing where the bar sits is the hard part. You decompose real work, take the standard from the practitioners who hold it, and encode it so an agent cannot shortcut it.
You will apply what Mercor has learned building benchmarks with domain experts, and devise new methods, so that evals and rubrics keep improving and so do the agents measured against them.
That work only scales with a platform behind it. This is a platform engineering role with good depth of understanding in evals. You will build the verifiers, the environments agents are measured in, and the grading infrastructure that runs at scale, abstracted across customers, domains, and tasks so that every run becomes evidence the next agent inherits instead of starting over.
Read more about how we think about this: Agent Eval Systems