Note: The job is a remote job and is open to candidates in USA. The client is building a standalone, market-ready AI software product within its AI and Digital organization. The AI Harness Engineer will design, build, and ship production-grade full-stack features while owning agent orchestration, tool integrations, context engineering, evaluation infrastructure, and telemetry. The role also involves deploying the product with customers and translating customer needs into software.
Responsibilities
- Design, build, and ship production features for a new AI product — hands-on, full-stack, and at high velocity, from first commit through production rollout
- Own real production engineering: hardening, scalability, reliability, observability, and security. Code reaches production through enterprise security review, and you are expected to navigate that, not route around it
- Architect the product to generalize across domains. Build a durable core product experience rather than rebuilding from scratch for every new domain or customer
- Move fluidly across the stack and across AI modalities — backend services, frontend surfaces, and use cases spanning language, vision, classification, and voice, depending on the business problem in front of you
- As the product ships, engage directly with customers: deploy it, integrate it, support adoption, and help it succeed in the field
- Translate customer pain into product, and explain clearly what you built, why the architecture is right, and what business problem it solves — through demos, narrative, and working software
- Build and own the agent loop: orchestration, state management, step limits, and termination conditions
- Design the tool layer — tool schemas, argument validation, dispatch, and error surfaces a model can actually recover from
- Own context engineering as versioned engineering artifacts rather than loose prompt files: retrieval, prompt assembly, compaction, memory
- Build evaluation infrastructure: golden datasets, task-specific rubrics, trace-based scoring, and regression suites that run in CI
- Instrument tracing and telemetry across every agent step, including token consumption, latency, and tool failure rates
- Build integrations around existing harnesses and tools rather than reinventing what already works
Skills
- Hands-on engineering ability, current: You write and ship production code today — not a resume that says you used to. We will ask for real, referenceable work that went to production
- Full-stack range: Python is required and universal on this team. Rust and Go are valuable. Frontend work centers on TypeScript and React, with some Angular — you should be able to move between them, using coding agents to close the gaps
- Full software development lifecycle: Demonstrated ownership from design through deployment: production rollout, scale, security, and observability. Not prototypes that never shipped
- Enterprise experience: You have shipped inside a large or regulated organization and know how to work the system — security reviews, compliance constraints, change control, and enterprise-grade product development. KPMG is heavily regulated, and this is where newcomers most often struggle
- AI-first in your own workflow: You use AI across your entire development process — coding, testing, QA, review — not just in the products you build. Fluency with Claude and Claude Code, Cursor, Copilot, Codex, Devin, agent frameworks, frontier models, and the open-source LLM ecosystem. You understand the loop around the model: prompting, tool use, evaluation, and how to get real leverage from these systems
- Conceptual depth in agentic AI: You can explain the difference between an agent harness and an agent loop, and speak concretely about generative AI, tool integration, orchestration, and evaluation
- Communication: You can explain an architectural decision and the business problem it solves to a non-obvious audience. Storytelling and customer-facing presence are innate, hard to coach, and explicitly screened for
- Product mindset and startup velocity: Comfortable with ambiguity, biased toward shipping, motivated by owning outcomes — while operating within enterprise expectations
- Work authorization: Authorized to work in the United States without current or future employer sponsorship
- A mix of startup and enterprise experience. Startup agility paired with enterprise discipline is the combination that works best here — experience at a known startup or product company, rather than solo or side projects alone
- Direct agent harness engineering experience. This is the most differentiated signal available, and an outdated framework is not a problem — the judgment picked up building matters more than the specific stack
- Experience in a regulated environment (financial services, healthcare, government, or similar)
- Familiarity with the Model Context Protocol or comparable tool integration standards
- Open-source contributions to agent frameworks, eval tooling, or developer infrastructure
- Experience taking AI products from experimentation into production deployment
Company Overview