Research Engineer

We are making AI that can build megaprojects: power plants, factories, data centers, and the physical infrastructure civilization runs on. Today this work requires large teams coordinating over long timescales. We are making agents that do that coordination directly.

The research that matters most to us comes straight out of deployment: getting agents to multi-task, learn, remember, and operate reliably over long stretches of time. You'll own that work end to end, building agents, training models, and designing the benchmarks, environments, and evals that tell us whether it's working. You'll set research direction with our founders, grounded in real deployment and test data, and translate what we learn into production.

Agents forget what they've committed to and what users teach them, so we're building long-horizon learning benchmarks (starting from our Horizon work) for memory once history overflows the context window. Real work also comes in bursts. Give an agent three tasks at once and it drops most of them, so we need environments for multi-tasking, stakeholder management, requirements-gathering, and scaled caution around irreversible actions. Underneath it all sit open problems we have to crack: temporal reasoning and planning, and a verifiable, evidence-backed model of the state of the world.

You've built agents, trained models, and created benchmarks, environments, or evals (ideally all of the above), and you can connect your own work to the long-term goal of getting AI to build in the real world. We're especially excited about current PhDs and post-docs in continual learning, test-time training, long-context architectures, and long-horizon tasks, but we care more about what you've shipped than your credentials.