Senior AI Agent Harness Engineer

NVIDIA

Town of Texas (WI)

On-site

USD 184,000 - 357,000

Full time

14 days+

Application generator

Stand out for this role — generate a tailored resume and cover letter in about a minute.

Get past ATS filters

Benefits offered by this job

Equity

Job summary

NVIDIA in Wisconsin seeks a Senior Software Engineer to build core libraries, evaluation systems, and reusable harness capabilities for AI agents. You will work across open-source agent harnesses, sandboxed execution, model-provider infrastructure, and agent benchmarks to improve safety, speed, and trust in agents.

You will own design decisions, diagnose failures from traces and logs, and contribute upstream with tests, documentation, and upstream projects.

Qualifications

Bachelor’s or Master’s degree in Computer Science, Artificial Intelligence, Applied Math, or related field, or equivalent experience.
8+ years of hands-on software-engineering experience, with ownership of production systems; architecture design leadership.
Expert Python skills and working proficiency in Rust, Go, C++, or TypeScript, with cross-language debugging ability.
Experience building or extending LLM agents, tool-use loops, model-provider integrations, developer tools, or evaluation infrastructure.
Solid understanding of asynchronous execution, subprocesses, containers, sandboxing, callbacks, retries, networking, and distributed compute.
Experience crafting reproducible experiments, benchmark methodology, performance investigations, regression gates, or reliability analysis under nondeterministic conditions.
Record of shipping open-source or developer-facing software with testing, API development, documentation, and code reviews.

Responsibilities

Define and evolve the technical approach for evaluating and improving open agent harnesses across frameworks, models, and benchmark environments.
Build and operate scalable, reproducible evaluation systems spanning CI, scheduled compute, sandboxed execution, artifact provenance, trace capture, and analysis.
Diagnose agent failures from traces, logs, tool results, and benchmark artifacts, separating harness, evaluator, environment, inference, and model causes.
Design controlled experiments and promotion criteria that distinguish genuine harness improvements from noise or model-specific gains.
Ship focused harness and runtime improvements, including upstream open-source contributions, detectors, regression tests, and maintainable documentation.
Partner with Relay, Hermes, model, infrastructure, and research teams to turn recurring optimization needs into reusable runtime, trace, and evaluation capabilities.