Senior Staff Software Engineer - Enterprise AI Platform

NVIDIA•Santa Clara, CA

1d•$200,000 - $322,000

About The Position

We're building the platform that lets long-running autonomous agents operate safely inside NVIDIA's enterprise. These are not assistants on a developer's laptop. They are fleets of agents deployed in the cloud, running continuously at scale on shared accelerated compute. They take on real work across enterprise systems, so people get far more done than they could before. This role defines the constructs that agents are built from: the blueprints they start from, the tools, skills, and plugins that power them against enterprise data, the runtime safety harness that keeps them in bounds, and the connections into credential management, sandbox, memory, and observability. The team designs and ships these building blocks so that agent developers across the company can stand up a new agent, wire it in, and run it for days or weeks. Security and safe execution come out of the box, not something each team has to get right on its own. Today an agent runs inside a single harness. Claude, Codex, and open-source agent harnesses each work differently underneath, with their own execution model, tool interface, and telemetry shape. The platform smooths over those differences, so a single skill, safety policy, or trace works the same no matter which harness is running. We want to enable agents that act on a person's behalf, governed and secure, continuously evaluated and self-improving. These agents coordinate and hand work off to each other, with identity and policy following every hop. They route and tune themselves across harnesses from live eval signals, and get better from their own production telemetry instead of waiting on a human to retrain them. Have you run agents on a harness like Claude or Codex and hit the walls that show up when they run for real, for days, against live systems — and wanted them to learn from it on their own? We're building the platform that solves those problems once, for every team.

Requirements

BS or MS in Computer Science, Engineering, or related field (or equivalent experience)
12+ years building distributed systems, infrastructure, or developer platforms at scale
Hands-on experience building agents on a harness, exposing them as APIs, and shipping them with CI/CD
Experience deploying and operating long-running services on container orchestration platforms
Experience with the building blocks of scalable systems: messaging, caching, and durable storage
Proficiency in Python, Go, Rust, or similar

Nice To Haves

Built a safety or policy engine that enforces rules on agent actions at runtime, with approval gates and kill switches
Designed evaluation and feedback loops for agent behavior, tied to versioned skills or blueprints. Built self-evolving loops where agents improve from their own eval and production signals, on the latest agent harnesses — the closed-loop, self-improving side you want to attract
Applied security fundamentals like threat modeling, authentication and authorization, least privilege, secrets management, and token exchange
Designed AI data platform components like ingestion pipelines, vector stores, and retrieval APIs
Shipped platform building blocks adopted by multiple engineering teams. Led complex technical projects like migrations or greenfield platform builds, aligning teams and writing clear design docs

Responsibilities

Design agent blueprints with clean interfaces for authorization, sandbox, memory, observability, and skills, so a new agent inherits its enterprise integrations from the platform.
Build the runtime safety harness: a policy engine that checks every action before it runs, rate and budget caps, circuit breakers, approval gates, action allow-lists, and a kill switch that works even when an agent goes rogue.
Enable composing and orchestrating agents: skills as first-class units with declarative manifests, multi-agent orchestration for delegation and handoff, and support for headless, long-running, autonomous agents.
Broker credentials so multi-agent systems can authenticate and authorize without ever touching secrets, with least-privilege scoping on every token.
Provide checkpoint and recovery so an agent resumes cleanly after a crash or restart.
Instrument observability and evaluation that span harnesses: decision-level traces with correlation IDs, what the agent saw and chose and why, cost anomaly alerts that catch looping, and quality scoring across skills, whole agents and products. Then close the loop: turn those signals into insights that make the agents better.
Keeping the platform healthy and reliable, including taking part in an on-call rotation.