Research Engineering Manager, AI Agent Harness (Product R&D)

𝗔𝗯𝘂 𝗗𝗵𝗮𝗯𝗶 (𝗨𝗔𝗘) · 𝗙𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝗧𝗲𝗮𝗺, 𝗥&𝗗 · 𝗘𝗠, 𝗟𝟱 · 𝗠𝗝𝗗𝟬𝟯.𝟭.𝟬Hands-on Research Engineering Manager to establish the R&D team that builds an 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝗵𝗮𝗿𝗻𝗲𝘀𝘀 and evaluation regime for 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗶𝗻𝗴 𝘄𝗲𝗯 𝗯𝗿𝗼𝘄𝘀𝗲𝗿𝘀 with open-weight LLMs. Greenfield.THE TECHNICAL CHALLENGEWe build agents on open-weight LLMs that operate web applications through the browser. The engineering work is the harness around the model: the abstraction that lets a pipeline or another agent drive a browser agent from the terminal, and the evaluation regime across levels of abstraction, from single actions ("find the element matching this description") to whole tasks, run against simulated applications before a live one. Agents maintain the evaluation suite; humans define its schema. Progress is measured against benchmarked results. You lead the team that builds this with the ability to write code.KEY RESPONSIBILITIES• Set the technical direction of the harness and the interfaces between its levels of abstraction• Establish the evaluation regime and the fortnightly score, and hold the team to measured results• Run the R&D team on a two-week rhythm: planning, review with working software, retrospective, written post-mortems for failed experiments and incidents• Coach, level and hire the engineers; interface with product management on the backlog and with the platform team on what the harness exposesDESIRED QUALIFICATIONS• Managed both a research-cadence team and a sprint-cadence team, and can articulate the interface between them• Built an agent harness or an evaluation harness for computer-use systems, agents or reinforcement learning, with an evaluation culture recognised outside the team, and can say which abstraction was right and which was wrong• Has taken apart the harnesses of coding agents (OpenHands, SWE-agent, Aider, DeepSeek Harness or comparable) and can explain how they work in plain terms• Has fine-tuned an LLM, from data preparation to evaluation on a held-out setEXPECTED QUALIFICATIONS• T-shaped: deep in one domain, with working breadth in a neighbouring one• Structures a large, incompletely specified problem and drives it to a working result independently• Managed a team building on LLMs or machine-learning models in production, with evaluation pipelines as a management requirement and a clear account of how quality was gated• Turns research or prototype code into reference implementations others reuse, using AI coding tools daily and verifying their output——HOW WE WORK• Product engineering: we own what we build and run it in production• Small teams, two-week cycles, working software at every review• AI coding tools are part of the standard workflowWHAT WE OFFER• Founding-team scope with a direct line to the Product CTO• A supported track to the UAE Golden Visa (ten-year residency) for AI and technology talent, and relocation support including the UAE AI Specialist visa• MacBook Pro M5 Max with 64 GB RAM or more, gear, Nvidia B200s and a serious budget for AI tokens, credits and tooling• Up to six weeks per year working remotely from anywherePROCESS• Introductory call with our recruiter• Technical conversation with the Product CTO• Practical session; the format is agreed with you• In exercises, AI tools are allowed and expected (no LeetCode)WHO WE ARENew product organisation backed by a semi-government in Abu Dhabi. International, ex-FAANG team. Completely greenfield, with a modern tech stack.——REQUIREMENTS TO BE CONSIDERED• Clear written and spoken English; has reported technical progress to non-technical executives on a fixed cadence• 6+ years in engineering, of which 2+ leading a team of engineers where work was accepted on measured results• Still coding, able to review TypeScript and Python; own code operated in production at a product company or a startup• Bachelor's degree in any field, or self-taught with a track record of open-source contributionsRELATED TECHNOLOGIES AND CONCEPTS• Agent harnesses and frameworks: OpenHands, SWE-agent, Aider, DeepSeek Harness, Hermes Agent, LangGraph, smolagents• Programmatic prompt and pipeline optimisation: DSPy, TextGrad, Ax• Browser automation and computer use: Playwright, Chrome DevTools Protocol, Stagehand, Browser Use, Steel Browser• Evaluation and observability: evaluation harnesses, LLM-as-judge, trajectory evaluation, Inspect, Langfuse, OpenTelemetry; benchmarks such as Online-Mind2Web, WebArena, OSWorld• Models and serving: open-weight LLMs, vLLM, SGLang• State and orchestration: state machines, event sourcing, XState, Trigger.dev, Temporal, Kubernetes, TypeScript, Python