Staff Software Engineer, Gustie Assist

About Gusto

At Gusto, we're on a mission to grow the small business economy. We handle the hard stuff — payroll, health insurance, 401(k)s, and HR — so owners can focus on their craft and their customers. With teams in Denver, San Francisco, and New York, we support more than 500,000 small businesses nationwide and are building a workplace that reflects the people we serve.

All full-time employees receive competitive base pay, benefits, and equity (RSUs) — because everyone who helps build Gusto should share in its success. Offer amounts are determined by role, level, and location. Learn more about ourTotal Rewards philosophy.

AI is a fundamental part of how work gets done at Gusto. We expect all team members to actively engage with AI tools relevant to their role and grow their fluency as the technology evolves. AI experience requirements vary by role and will be assessed during the interview process.

Gustie Assist builds Otto, the AI agent that Gusto's support advocates rely on to solve customers' hardest payroll, benefits, and compliance problems. When a small business hits a wall and reaches out, the advocate turns to Otto: it reasons over the customer's full context, answers in the moment, and increasingly takes real actions across Gusto's systems on the advocate's behalf to make the customer whole. We're a small platform team building an agent that reasons, acts, and has to get it right when the stakes are someone's payroll.

This is unusually broad, hands-on applied-AI work: you'll own an LLM agent runtime and the full stack around it, including the tool layer that lets it act over Gusto's systems, the Rails/GraphQL backend, the advocate-facing frontend, and the eval system that keeps it trustworthy while it touches live customer accounts. We keep the path from idea to production short, and measure ourselves on customer outcomes, not model novelty.

Here's what you'll do day-to-day:

• Design and build across the Otto platform, including the runtime and agent graph, the tool/action layer, the Rails/GraphQL backend, and the advocate-facing experience, owning a meaningful slice end-to-end
• Expand what the agent can safely do: autonomous, tool-using actions with human-in-the-loop approval, built on a secure, auditable identity and permissions model
• Build the eval and observability systems (LLM-as-judge + deterministic scorers, offline gates and online production monitoring) that let us change the agent in production weekly without regressing correctness or escalation behavior
• Work directly with LLMs, including prompting, tool/function design, retrieval, and the feedback loops that turn real advocat