The hard part of AI research isn't just making something work.
It's figuring out whether it's worth building, who needs it, and what it takes to make it work beyond the demo. We're a cross-functional team in the AI Platform Research pillar, working at the front edge of AI research, where the technology changes faster than the product can be specified.
We work alongside researchers, not after the research is done. We identify the assumptions behind an idea, test the riskiest ones early, and turn what we learn into better decisions about where the work should go next — practitioners as much as evaluators, building and iterating on real systems while the questions are still open.
Right now that means DAM, our open source platform for running long-lived AI agents, and the inference and serving systems underneath it. The projects will change; the questions — what agents need from a platform, what makes agentic work practical to run — probably won't.
How we work
Learn before we build
The goal isn't to build as fast as possible. It's to find out as fast as possible whether something is worth building.
Start small
Build just enough to answer the question in front of us, then stop or change direction once we have the answer.
Stay close to users
Decisions come from real needs and real use, not from guesses about what people might want.
What we do
Agent infrastructure
DAM, our open source platform for running long-lived AI agents — each in an isolated sandbox behind a zero-trust credential gateway, driven from the browser, the CLI, Slack, or on schedules. We're still exploring where the line falls between what the agent handles and what the platform should.
Inference at scale
The serving substrate agentic research runs on, tuned for the throughput and cost of long-running autonomous work — so experiments are bounded by ideas, not by capacity.
Research validation
Interviews, studies, flow boards, coded prototypes — whatever makes an idea concrete enough to put in front of researchers and challenge before it's real. What breaks tells us what to build, change, or stop.