About
Agents only learn what the environment rewards. If the tasks are too easy, too narrow or too easy to game, the agent learns the wrong thing and the benchmark still goes up. This role exists to build the layer that prevents that: the environments, tasks and reward signals agents train against.
This is a well-funded frontier AI company building computer-use agents and the models behind them. The task library already spans hundreds of apps and tens of thousands of tasks, across GUI, API, MCP and coding. Every extra task has to be realistic, verifiable and hard enough to matter.
You’d be joining the team building the playground itself: synthetic websites, structured workflows, task sets, and evaluation environments where agents can act, fail, retry, and learn.
What you'll do
- Build training and evaluation environments for agents, from synthetic websites and workflows to full app simulations
- Design task suites and success criteria that reflect useful real-world work
- Define reward signals and verifiers that are hard to game
- Generate synthetic data at scale and keep its quality measurable
- Turn documentation, tools and existing workflows into interactive agent tasks
- Take agent failures from research and product and turn them into better environments
- Build the tooling to generate, run and measure large task sets reliably
What you'll need
- Strong software engineering in Python, ideally with web or backend systems
- Experience building simulations, evaluation systems, internal tools or synthetic environments
- Ability to turn ambiguous workflows into tasks with measurable outcomes
- Good instinct for what makes an environment realistic and useful for agents
- Comfort between engineering, research and experimentation
- High ownership and a practical mindset
Optional Bonus
- Browser automation, agent frameworks or evaluation harnesses
- Reward design or verifiable-reward task setups
- Synthetic data generation for LLM training
Shortlisted candidates will be contacted within 48 hours.
