Prompt Engineer, Multi-Agent Systems

You want to work on a system of agents that work together. Some of them talk to customers, and many of them never say a word to anyone. They interpret, decide, and pass work to each other, and you understand that the quality of the whole system depends on every link in that chain holding up in production. Making that behave reliably, at scale, with real customers on the other end, is the problem you want to own.

You treat prompts as engineering artifacts, not text you tweak until a demo works. If you've ever shipped a prompt change that fixed one case and quietly broke nine others, you know why the evaluation layer matters as much as the prompt itself. And if you get a small thrill from a regression suite going green after a nasty bug, you'll feel at home here.

You'll also recognize how we work. We measure what we ship: changes come with evidence behind them, a dataset, a score, a clear before and after. We ship weekly, so ideas get tested against real production behavior quickly rather than studied at length. Craft matters, and so does rigor — getting an agent to sound right is satisfying, but proving it still sounds right across thousands of conversations is the actual work.

The surface area grows every quarter with new agents, new channels, and new products. You'll own a meaningful slice of it from day one.

Design prompts for agents that work together. Define what each agent knows, what it's allowed to decide, what it hands off, and what it must never do, including the contracts between agents that keep the system coherent end to end.

Build the evaluation layer. Every prompt you ship comes with a dataset and a regression suite. You'll curate test cases from real production behavior, write evaluator prompts, define scoring rubrics that actually track business outcomes, and catch regressions before customers do.

Make the system observable. Tracing, prompt versioning, datasets, experiments, scoring. You'll instrument what isn't yet instrumented and use the data to find where agents are actually struggling, which is often somewhere other than where we'd guess.