ABOUT THE COMPANY

We're building autonomous research agents for recursive self-improvement (multi-agent systems that propose, run, and analyze machine learning experiments). We're a small, well-funded team that works onsite in San Francisco.

ABOUT THE ROLE

You'll be researching the agents at the core of our work: multi-agent systems that conduct automated machine learning research and discovery. You'll design how these agents plan, decompose problems, choose what to try next, evaluate their own outputs, and recover from mistakes.

This is a deeply open-ended research role. The benchmarks for agents that do real research don't exist yet, and inventing them is part of the job. You'll move between method design, careful experimentation, building evaluation frameworks, and shipping into production. Real autonomy, real ownership, and the corresponding responsibility for choosing well.

WHAT YOU'LL DO

Design methods that improve how our agents plan, decompose tasks, use tools, manage context, and recover from failures across long-horizon research workflows

Develop multi-agent coordination patterns: how multiple agents share context, divide labor, supervise each other, and combine their outputs

Build and maintain evaluation frameworks for agent capability on open-ended tasks (the kind where the right answer isn't pre-specified)

Run rigorous experiments to characterize what works, what doesn't, and why: controls, ablations, statistical significance

Co-design agent architectures with engineering teammates; ship the most promising methods into production

Read deeply across the agentic ML, planning, RL, and tool-use literature; bring useful work from outside in

Share findings internally so the rest of the team builds on them

Help shape research direction across the team: agentic research taste compounds when discussed openly

WHAT WE'RE LOOKING FOR

Strong track record of ML research with focus on agents, RL, LLMs, planning, tool use, or multi-agent systems

5+ years of hands-on research experience in industry or academia

Comfort designing experiments and running them end-to-end at scale

Track record of building evaluation frameworks for capabilities that aren't easily benchmarked

Bias toward shipping research, not handing it off

Strong written communication: you can compress a result into a paragraph that changes what someone else does next

Comfort with ambiguity: open-ended problems without fixed benchmarks are the work, not a frustration

Published research at NeurIPS, ICML, ICLR, COLM, RLC, or comparable venues

NICE TO HAVE

PhD in ML, statistics, CS, or adjacent

Published research on agentic systems, tool use, long-horizon planning,

multi-agent coordination, or self-improvement methods

Open-source contributions in the agentic ML ecosystem (coding agents,

research assistants, autonomous workflows)

Experience with reasoning models, chain-of-thought / scratchpad methods,

or supervised fine-tuning for agentic behaviors

Background in evaluation methodology for capabilities that don't have

established benchmarks

THIS ROLE IS PROBABLY NOT FOR YOU IF

You want to focus on a single stable benchmark: our agents work on open-ended problems and the targets shift

You prefer to keep research paper-only; these agents need to actually work- You'd rather work alone than share research taste openly with a small team