AI Engineer

About the role

The Agent reads an organization's records, uses the product's tools, prepares drafts and waits for a person to approve them. It acts with the access of the person it helps and no more, and by default nothing it prepares changes or goes out until someone says yes.

That line is the product. You work on everything inside it: how the Agent plans a job, which tools it has, what it says when it is done, and how we know it got better and not just different.

You work in TypeScript, in the same codebase as the product, with Next.js, PostgreSQL and a job queue under it.

What you would do

• Improve how the Agent plans and carries out a job: choosing tools, handling errors and refusals, and ending on something finished.
• Add and shape the tools the Agent uses, so that each is typed, permissioned and clear to a model, and reads only what the person it helps may read.
• Write evaluations for the Agent: scenarios, graders and regression sets that show whether a change made it better or worse.
• Decide with numbers when a prompt, a model or a tool change is better, and say when it is only different.
• Protect the approval line: test that what waits for a person still waits, in every mode we ship.
• Read real runs, the failures first, and turn each failure into a test.

What we look for

• You have built software that people used, including a feature on a language model, and you measured whether it worked.
• Strong TypeScript: the code is TypeScript throughout.
• You test model behavior the way you test code: with cases, expected results and a record of what changed.
• You know when a model should ask a person, and you design for it.
• You keep the cost and the time of each Agent run in view.
• You write clearly: tool descriptions and prompts are text that other people maintain.