Mercor Product Manager, Environments at Mercor to own the platform creating training tasks for frontier AI agents. Define requirements, set quality standards, and scale task generation.
Responsibilities
As a Product Manager for Environments, you will own the platform that turns a world into thousands of hard, verifiable training tasks. You will decide what we build to push agents past their current limits, set the quality bar that labs hold us to, and turn work built for one project into products every team can use. What you ship helps shape what goes into frontier training runs, and what the next generation of agents can do.
Frontier Capability Gaps: Work with researchers at frontier labs to find where their agents fall short, and turn those gaps into clear requirements for the worlds, tasks, and verifiers we build next.
Task Generation at Scale: Build task generation systems that turn a single world into thousands of realistic, hard, verifiable tasks, with controls that let teams steer what gets produced.
Environment Quality Bar: Set the bar for when a world is ready for tasking and when a task is difficult, solvable, and fairly graded, and build the checks that hold every project to it.
Run-by-Run Feedback Loops: Build the feedback loops that show how every run is performing, so teams can find the bottleneck and push yield and difficulty higher with each iteration.
Shared Environment Platform: Turn what works on one project into a platform every team can use, including the tools experts use to review and improve tasks, so each new environment is faster to build than the last.
Roadmap and Success Criteria: Define success in terms of high-quality tasks delivered, balancing throughput, difficulty, and cost, and use lab feedback and agent failures to decide what the platform builds next.
You might thrive in this role if you:
Experience: Have 3+ years as a product manager, prior experience in SWE, ML, or DS preferred.
High Ownership: Take full accountability for outcomes, not just outputs.
Independence: Are able to self-direct in ambiguous contexts, creating clarity for others.
Technical Fluency: Are comfortable building scrappy MVPs yourself when needed to validate direction, unblock engineering, or accelerate iteration.
