Staff ML Engineer, Agent Training & Environments

Labelbox•San Francisco, CA

13d•Hybrid

About The Position

Labelbox is the RL data factory for advancing frontier agent capabilities. We build the data, environments, and evaluations that frontier labs use to train and judge their agents. This role sits where training meets infrastructure. You will run the experiments and build the systems that run them: environments agents act in, verifiers that decide whether they succeeded, and the fine-tuning pipelines that turn that signal into a better model. We're looking for someone who does both halves — the engineering throughput of a strong platform engineer, and real depth in post-training agents. The bar is high: engineers with strong judgment who set technical direction, turn prototypes into reliable systems fast, and are at the frontier of agent-first engineering practice.

Requirements

A 3+ year track record of shipping systems that customers and other engineers still rely on. Exceptional throughput, without the quality tax. You ship a lot, you review a lot, and the v1 you ship becomes the foundation the rest of the team builds on.
Strong system and API design judgment. Hard architecture calls land with you: you make them, defend them under pressure, and update fast when someone else is right.
You ship production code with coding agents daily. You know where they break and what it takes to make them reliable, and you use that to move the whole team faster.
You build the substrate other people's work runs on — tooling, CI, harnesses, libraries — and you treat that as the job, not a distraction from it.
You move fast in ambiguous, startup-pace environments, with influence over authority.
Deep proficiency in Python, and comfort across the rest of the stack.
You have fine-tuned models for agentic tasks and made them measurably better. SFT plus at least one RL method (GRPO, PPO, DPO, or similar) in production.
You have built environments agents operate in, and you know why reward and task design is where most of the difficulty actually lives.
You have designed verifiers or graders for open-ended work, and you know how they get gamed.
You debug training runs forensically and methodically.
You reason about compute-economics. You know what an experiment costs, when a run is not worth finishing, and how to get the same signal for a tenth of the spend.
You write up what you learned so it changes what the team does next.

Nice To Haves

Experience with agent harnesses and coding agents as subjects of training and evaluation.
Multi-tenancy and isolation for untrusted agent execution: sandboxing, egress control, credential handling.
Background in production distributed systems, ML infrastructure, or data systems at scale.
Experience working directly with frontier labs or other highly technical customers.

Responsibilities

RL environments for agentic tasks: task definitions, tool surfaces, state and reset semantics, reward design — and the harness that runs thousands of them in parallel.
Verifiers and graders: programmatic checks, LLM judges, rubric pipelines, pass@k scoring. Deciding what "the agent succeeded" means, and making that judgment trustworthy at scale.
Fine-tuning pipelines that turn evaluation signals into measurable agent improvements — SFT and RL, from data collection through training to checkpoint evaluation.
Eval systems that run millions of agent trajectories to measure model and product quality.
Training and serving infrastructure that scales to the throughput frontier labs need: multi-launcher orchestration, long-running job fault tolerance, cost accounting.