Agent Evaluation Infrastructure Engineer

San Francisco, California | Primarily On-site

We are seeking an Agent Evaluation Infrastructure Engineer to build the environments, evaluation systems, and supporting infrastructure used to train and assess long-horizon enterprise AI agents.

The Opportunity

You will work on the engineering and research problems behind realistic agent environments, post-training systems, and reliable evaluation of complex multi-step workflows.

Key Responsibilities

• Design evaluation environments for long-horizon enterprise agent workflows.
• Define tasks, state, tools, graders, and reward signals used to evaluate and improve agents.
• Build high-fidelity representations of complex enterprise software environments.
• Develop infrastructure for rollouts, orchestration, trajectory inspection, and grader pipelines.
• Measure both correctness and efficiency across multi-step agent behavior.
• Investigate evaluation failures, reward-quality issues, and agent behavior.
• Build production-quality systems rather than notebook-only research prototypes.

Required Qualifications

• Hands-on experience with AI environments, evaluations, reinforcement learning infrastructure, or related agent-training systems.
• Strong software engineering fundamentals.
• Demonstrated ability to build and ship technical infrastructure.
• Understanding of evaluation methodology, reward design, graders, and agent trajectories.
• Ability to work across languages and technology stacks based on system requirements.

Candidate Profile

A PhD is not required. Strong engineering and shipped environment or evaluation systems are more important than academic credentials or publication history.

Seniority

The opportunity is open to exceptional new graduates, early-career engineers, and experienced senior candidates. Selection is based primarily on engineering strength and relevant technical work.

Work Arrangement

The role is anchored in San Francisco with a strong preference for in-person collaboration. Limited flexibility may be considered case by case for exceptional candidates.