Senior Staff Software Engineer

The team

Build Agent is ServiceNow's AI coding assistant, purpose-built for the platform's metadata-driven substrate - operating natively across ServiceNow's scoped applications, tables, and metadata types. This role leads the team responsible for the Build Agent evaluation framework, model support, and telemetry.

The framework is critical to Build Agent success — it's already driven measurable wins: recovering Build Agent correctness, cutting latency and inference cost ahead of a major release, and catching high-severity defects that manual testing missed before they reached production. The team is 8 engineers and is currently the bottleneck on its own scale — this role exists to convert a high-performing initiative into a durable, scalable function.

Key areas of focus include:

• Evaluation Infrastructure: Golden prompt sets, Pass@1 and functional scoring, failure categorization (plumbing vs. metadata-creation failures), and coverage across ServiceNow metadata types and UI workflows.
• Model Support & Benchmarking: Structured evaluation of candidate foundation models against the production default, with failure-consistency analysis to separate scaffold/tuning issues from genuine capability gaps.
• Telemetry: Token usage, cache efficiency, and inference cost tracked alongside correctness as first-class release signals.
• Cross-Team Arbitration: Prioritizing eval coverage, defining what "good" means across surfaces owned by different contributing teams, and turning eval signal into committed fixes by the owning teams rather than open-ended findings.
What you get to do in this role

• Own eval strategy and roadmap: prompt set design, scoring methodology, failure-mode taxonomy, and telemetry instrumentation.
• Coordinate across the multiple teams contributing to Build Agent to arbitrate quality standards and drive eval findings to committed, owned fixes — not just reports.
• Lead model benchmarking efforts, making data-backed recommendations on model support decisions (e.g., evaluating frontier coding models against the incumbent default).
• Solve ambiguous, cross-cutting problems where eval signal, model capability, and platform architecture intersect — and where "correct" isn't obvious until you've defined the bar.

Department: Engineering, Infrastructure and Operations
Function: Engineering
Experience Level: Not Applicable