Harness Research Engineer, Agents

36Labs; Greater London; Hybrid; GBP 90,000 - 160,000; Full time

Job summary

36 Labs is developing an autonomous agent harness that plans, generates, and refines content, learning from audience engagement data. You’ll own the production code and drive implementation from day one, collaborating with the CTO, backend engineers, and content engineers to shape the system.

You will lead design decisions, set technical direction, and grow into technical leadership as you scale the harness across experiments and real user workloads.

Qualifications

• 4+ years in ML engineering, with end-to-end ownership of systems.
• Strong Python and production-grade software engineering practices.
• Experience shipping an agent system with models calling tools.
• Ability to design evals with rubrics and model judges.
• Experience running benchmarks and A/B tests.
• Mentored junior engineers and led technical work.

Responsibilities

• Design and build the creative harness end to end.
• Create sandboxes and memory layers for agents.
• Own evals with datasets, rubrics, and human labels.
• Build benchmarks to measure changes across runs.
• Run experiments and A/B tests on prompts and tools.
• Turn performance data into signals for learning.
• Choose and adapt models per step; frontier or open-weight.
• Follow research and bring new techniques.
• Trace failure modes across full agent loop.
• Set ML engineering standards: reproducibility and docs.

Skills

Python

ML engineering

Production code

Code review

Team leadership

Education

BSc/MSc/PhD

CS/Math/Stats/Physics

Tools

PostgreSQL

TypeScript

LLM APIs

Open-weight models

Job description

36 Labs is a seed-stage startup that started out of the UCL Machine Learning lab. We bootstrapped to 10 people, and now we're on the next leg of growth.

AI can generate almost any form of media, but it still can't judge what captures human attention. AI-generated content on social platforms is growing fast, every creative workflow will be agent-assisted or fully agentic. Our work is improving how models reason about creative decisions, learning from how audiences respond to content.

The role

You'll lead engineering on our creative harness: the agent system that plans, generates and refines content, and learns from audience engagement data.

This is a hands-on engineering role. You'll write and ship the production code yourself, not hand off analysis or notebooks for someone else to build.

You'll have the autonomy to set how we build, measure and improve our agents, with scope to grow into technical leadership. You'll work closely with the CTO on direction, with backend engineers on the infrastructure the harness runs on, and with content engineers who use it every day.

What you'll do

• Design and build the creative harness end to end: planning, tool use, context, memory and multi-step execution, so agents can run long, open-ended work on their own.
• Build the sandboxes and environments our agents act in, and new memory layers that let them carry what they learn across long runs.
• Own our evals: decide what "good" means for creative work with no single right answer, then build the datasets, offline replays of real runs, rubrics, human labels and model judges (checked against people) that show when the harness gets better or worse.
• Build benchmarks for the harness, so a change to prompts, models, tools or agent design is measured over repeat runs, not judged from a few examples.
• Run experiments and A/B tests on prompts, models, tools and agent design, approach the results with statistical rigour, and ship what wins to production.
• Close the loop between audience response and agent behaviour: turn performance data into the signals, features and training data the harness learns from.
• Choose and adapt models for each step: frontier APIs, open-weight models, and fine-tuning where it beats prompting on quality, cost or latency.
• Follow the research closely and bring new agent techniques and architectures into the harness to push what our models can do.
• Trace failure modes across the full agent loop, from planning to tool calls to final output.
• Set the standard for ML engineering: experiment design, reproducibility, documentation, and guidance for other engineers building the harness.

What we're looking for

• BSc, MSc or PhD in Computer Science, Mathematics, Statistics, Physics or a related field, or the equivalent in the form of experience. Strong Python.
• 4+ years in ML engineering, writing production code and owning the systems you built end to end. A data science background is welcome if you shipped what you built.
• Strong software engineering: production Python, testing, code review, and deploying and operating services.
• Built and shipped an agent system that real users or workloads depend on: models calling tools, planning across steps and acting on the results. A chat interface over retrieval is not enough on its own.
• Strong foundation in statistics, experimental design and research: you can tell a statistically significant improvement from noise.
• Designed and shipped evals for outputs that can't be checked automatically (creative, writing or open-ended agent tasks): turned fuzzy quality questions into rubrics, labels and validated model judges that drove real decisions.
• Built or run benchmarks for agents or harnesses, and can explain how you knew a change was real and not noise.
• Built autonomous agents that run long tasks unattended: sandboxed execution, long-horizon context and memory beyond retrieval. Working on new memory approaches is a big plus.
• Follow research closely, especially on evals and agents, and have put new methods into practice: you can name a recent paper and what you tried from it.
• Led technical work: set direction, reviewed designs, and have mentored more junior colleagues.
• Systems or projects you've built and can discuss in technical depth.

Nice to have

• Post-training experience: SFT, RLVR, RLHF, preference optimisation or RL for agent behaviour.
• Ranking, recommendation or causal inference work, especially predicting engagement.
• Multimodal models: image, video or audio generation and understanding.
• Founding or early engineer at a startup, or equivalent ownership inside a larger company.

Our stack

Python, TypeScript, Postgres, LLM and generative media APIs, open-weight models.

We encourage you to apply even if your experience doesn't match every line above.