Member of Technical Staff – Research Engineering

Hone•San Francisco, CA

4d•Onsite

About The Position

Hone builds agentic AI systems, called engines, that work toward goals over weeks and months, deploy themselves, maintain their agents and code, involve people when judgment is needed, and improve over time. The company says it measures its work by economic impact and deploys engines across areas including revenue, procurement, risk, product, and operations. In this role, you will own how engines improve, including offline evaluations, simulations, memory, and self-improvement loops that turn production feedback into verified changes to skills, memory, code, and policies. The work focuses on engines operating within real companies, where goals may be underspecified and feedback is delayed, partial, and human-provided. The position is based in San Francisco and is in person, Sunday through Friday. Compensation is a competitive salary plus meaningful equity. Hone sponsors visas and supports employees through the immigration process.

Requirements

At least 2 years of engineering experience, or equivalent depth from research, open source, or personal projects.
Experience at a fast-growing startup.
Experience shipping applied ML systems from scratch as an engineer, not only publishing papers.
Experience running rigorous evaluations and experiments against production systems.
Strong fundamentals in ML evaluation, experimental design, statistics, and LLM agents.
Use coding agents every day while understanding every line of code shipped.
Build and tinker for fun, with genuine interest in how the technology works.

Nice To Haves

Harness, post-training, or applied AI experience at a lab or applied AI team.
Published research, especially on agent memory or continual learning.

Responsibilities

Build simulations from company context, including simulated users and connectors, to evaluate engines on tasks spanning weeks or months.
Turn production traces into simulations that can be rewound, forked with a change, and run forward while keeping the simulated world consistent.
Develop memory systems that store, retrieve, and compact what engines learn.
Build self-improvement loops that convert traces into verified changes to skills, memory, code, and policies.
Develop grading and credit-assignment methods for month-long trajectories.