Staff Machine Learning Engineer

Company Description

It all started when engineer Fred Luddy wrote code that automated a tedious task for his coworker, Phyllis. She cried tears of joy. That moment inspired Fred to build a company that could do that for everyone—freeing people from busywork so they could focus on meaningful work. Today, ServiceNow is the AI control tower for business reinvention. Our ServiceNow AI platform brings together any AI, any data, and any workflow— helping 85% of the Fortune 500® work smarter, faster, and better. We're building an AI-native culture where technology and talent are unstoppable together. And we're just getting started.

Join us to put AI to work for people.

Job Description

Team Overview

We build the AI layer of our CPQ platform — a set of Python services that let users configure, quote, and manage transactions through natural language instead of forms. This isn't a thin LLM wrapper. We're running multiple production agent architectures concurrently (ReAct-style tool-calling agents, hand-rolled LangGraph state machines, and the Harness — our from-scratch, industry-leading agent execution runtime). Our systems are backed by a first-party MCP surface into admin/product/rules/transaction systems and interoperate with other AI agents over the A2A protocol. Below that sits a conventional Java/Spring Boot microservices fleet and a React/TypeScript + Lit frontend that the agents ultimately drive.

Role Overview

We're looking for someone who already operates at a Senior-Staff bar in the agentic/LLM domain but is building out breadth across the rest of the stack. You'll be one of the most senior technical voices on how agentic systems get designed here — state management, tool boundaries, streaming protocols, prompt/context architecture, and multi-agent coordination — while staying credible end-to-end: able to read a Spring Boot service, unblock a frontend integration, or reason about a classical ML model pipeline when the problem calls for it.

What you get in this role:

• Multi-agent orchestration — LangGraph/LangChain agents over frontier LLMs for transaction editing, conversational configuration, and multi-product quote planning with plan/approve/refine loops and parallel task execution
• The Harness — we're crystallizing our own agent execution runtime into an industry-leading, state-of-the-art harness. Full-duplex sessions where a user can interrupt, redirect, or answer a clarifying question mid-execution while other work keeps streaming, built on a from-scratch async runtime rather than a bolted-on wrapper around someone else's agent loop. This is as much a performance and UX problem as a backend one — low-latency streaming, backpressure, live progress, partial results, graceful cancellation — and it's the part of the stack we're most invested in owning outright. You'd be a primary owner of where this goes next.
• MCP as a secondary interface — we maintain a first-party MCP server and clients into our admin/product/rules/transaction systems, but as the Harness matures it becomes the primary way our own agents interact with the platform, with MCP kept as the secondary, standards-based surface for external interop. You'd help decide what stays MCP-first and what moves onto the Harness.
• A2A protocol — agent-to-agent task delegation and streaming, surfaced through an external gateway so other systems (including core ServiceNow) can drive our agents directly
• Forward Deployed Engineering — expect real time embedded with customer- and product-facing teams against live deployments. Adapting the Harness and our agents to actual customer workflows under real constraints, not just building platform capability in the abstract
• RAG / context engineering — tenant-uploaded document ingestion, categorization, and aggregation into agent context. Prefix-cacheable prompt design for cost/latency
• Classical ML, when the problem isn't a good fit for an LLM — we have a separate PyTorch/scikit-learn training and serving pipeline (field-value prediction) that a whole-stack ML engineer should be able to read, extend, or evaluate against LLM-based alternatives
• Full-stack fluency — enough comfort in Spring Boot/Java services and the React/TypeScript + Lit frontend to unblock an integration end-to-end without waiting on a handoff

Qualifications

To be successful in this role you have:

• 8+ years building production software, with several years specifically shipping LLM-powered / agentic systems (not just API wrapper calls — real tool-use loops, state management, multi-turn orchestration)
• Experience in leveraging or critically thinking about how to integrate AI into work processes, decision-making, or problem-solving. This may include using AI-powered tools, automating workflows, analyzing AI-driven insights, or ex