Software Engineer, Agent Harness & Evaluation: Python, TypeScript, LLM Agents & Evals | San Jose, CA (On-site) | Up to $350K + Equity

"Listen to the technology; find out what it's telling you." — Carver Mead, the Caltech professor who taught a generation of engineers how to design a chip.

Sixty years of silicon progress, and the machines still can't tell us where they hurt. A droid in a film gets to whistle indignantly when something is broken. Real hardware just fails a test at 3am and leaves an engineer scrolling waveforms until sunrise.

That is the gap this role sits in. Our client builds AI agents that work on hardware engineering problems — the design, debug and verification grind that eats the schedule of every chip team on earth. The agents are already in production with paying customers. Your job is to make them genuinely good.

The short version

You will find out where the agents fail, design the evaluations that prove it, build the capabilities that fix it, and ship the result as production software people depend on. It is equal parts detective work and engineering craft.

If you have ever stared at an agent trace at 11pm thinking "that is not a model problem, that is a context problem" — and been right — you will recognise this job immediately.

What you will actually be doing

Breaking the agents on purpose. Finding the failure modes, then building evaluations that measure performance, reliability and efficiency honestly. Not vibes, and not a demo that works once.

Building new agent capability. Tool interfaces, context management, memory, and multi-step or multi-agent orchestration. The harness itself — the thing that decides whether a clever model becomes a useful product.

Reading traces like an investigator. Digging through experiment results for the real bottleneck, and knowing the difference between a genuine improvement and benchmark noise. That distinction matters more here than almost anywhere else.

Sitting with customers. Hardware engineers and domain experts will describe what is broken in their own language. You translate, build the fix, and own it end to end, typically inside a three-month window.

Turning experiments into software. Research that works in a notebook is a starting point, not a deliverable. You make it maintainable, tested and production grade.

You would be joining a sizeable and fast-growing group of software engineers, researchers, hardware specialists and field teams, working across customer discovery, evaluation design, applied research, product development and deployment. The work spans a wide range of hardware engineering problems, so no two months look quite the same.

The scorecard

This is a precise brief, so here is the hiring bar in full. Judge your own fit against it before you spend time on an application.

Dealbreakers

Hard skills — Real depth across LLM agents, orchestration, tool interfaces, memory, tracing, ML inference and transformer architectures. Systems you have run at scale, with large data and complicated tooling. Recommender systems, LLM gateways and infrastructure plumbing, side projects and straightforward RAG chatbots are not clearing this bar. Agents have to be the day job.

Soft skills — The ability to sit in front of customers, understand what they actually need, and deliver the technical answer yourself. This is scored as a dealbreaker, not a bonus.

Baseline

Seniority — 8 to 15 years as a software engineer, with serious time on large-scale agent systems. The bar is staff-plus. Strong engineers at the four to seven year mark are being passed over for this particular opening.

Work experience — Time at a rigorous technology company (think Applied Intuition, Palantir, Tesla) or at a serious AI startup (Fireworks, Harvey, strong Y Combinator companies).

Education — A CS, EE or similar degree from a top university.

Hard skills — Production software engineering in Python and TypeScript. Both, properly.

Soft skills — Self-directed and ambitious, with strong ownership, creativity and product judgement. The organisation is flat and people move fluidly between initiatives. Nobody will build your roadmap for you.

Nice to have

Experience or interest in electronic design automation (EDA), semiconductor domains, RTL, or hardware-software systems. You do not need it, but it is the single clearest differentiator in this pipeline. If you have worked at a major EDA vendor, or done on-device ML for consumer silicon, lead with it.

What will stop you

Work rights — US citizens and Green Card holders only. There is no sponsorship and no visa transfer on this role, so O-1, OPT and H-1B are non-starters.

Tenure — Under a year in your current role is flagged as a trait to avoid.

Why this one is worth the conversation

Most AI companies are still explaining what their product will eventually do. This one has paying customers among the largest semiconductor companies in the world, revenue to match, and a growth curve that has taken headcount up several times over in a year. The funding is substantial and the investor list includes names from inside the chip industry itself, which tells you something about whether the technology works.

The founding story runs through academic research on AI agents that predates most of the current wave. This team was working on agents before it was a category.

And the problem is hard in a way that will keep you interested. Chip design and verification is one of the most unforgiving engineering domains there is. An agent that is 95 per cent right is useless. Getting from impressive demo to dependable tool is the whole game, and it is the game you would be playing every day.

There is also unusual latitude in how you spend your time. The structure is flat, there are no silos between research and product, and scope is shaped around your strengths and interests. If you want to live mostly in evaluation and measurement, that is a job here. If you want to be in front of customers translating messy hardware problems into agent capability, that is a job here too. Most people end up doing some of each, and reporting runs to either the Head of Engineering or the Head of Research rather than through layers.

Compensation

Base: $250,000 – $350,000, with negotiating room for the right person.

Equity: Competitive, on top of base.

Location: San Jose, California, five days on-site. Relocation support is available.

Work rights: US citizens and Green Card holders only.

How to apply

Send your resume to [email protected] with a short note on an agent system you have shipped to production: what you owned across tools, memory and orchestration, and how you knew it was actually working. Include your salary expectation and confirmation that you are a US citizen or Green Card holder and able to be in San Jose five days a week.

If you are close but not an exact match on the scorecard, send it anyway and tell me where you think the gap is. I would rather have that conversation than guess.