Staff Research Engineer, Multi-Agent Scaling

EngineersOfAI; New York (NY); On-site; USD 170,000 - 210,000; Full time

Job summary

Anthropic is seeking a generalist engineer at the boundary of research and engineering. You will design and run large experiments, build the systems they run on, and dig into surprising results. The role involves moving from vague questions to running experiments quickly and scaling to very large agent teams.

This position emphasizes rigorous analysis of data, scaling performance and efficiency with team size and task horizon, and building trustworthy evaluations and tooling for researchers.

Qualifications

• Significant software engineering, ML or research engineering experience.
• Owned something substantial end-to-end such as a large system, an evaluation or benchmark, an agent product, or a research project.
• Genuinely enjoy both research and engineering work.
• Think quantitatively about complex systems and verify numbers before trust.

Responsibilities

• Design, run and interpret large-scale experiments on agent teams, reasoning about data.
• Investigate performance and efficiency as team size, compute and horizon grow, identify bottlenecks.
• Build and scale systems that run large agent teams reliably and debug scale-related failures.
• Design evaluations for long-horizon problems and ensure results are trustworthy.
• Develop tooling and metrics to help researchers see what a large agent team is doing and why.
• Collaborate with research teams to run their experiments on the platform and communicate findings.

Skills

Software engineering

ML

Research engineering

Job description

About Anthropic

Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.

About the role:

Large teams of agents are starting to solve problems no single agent can, from rewriting major codebases to formalizing landmark mathematics. Our team studies how these teams scale: what happens to performance, cost and coordination as the number of agents, the compute budget and the length of the task grow, and what has to change to keep getting returns from that scale. We build the platform and evaluations Anthropic uses to run and measure large agent teams, and other research teams build on them.

This role lives at the boundary between research and engineering. It is a generalist role on a small team: you'll design and run large experiments, build the systems they run on, and get to the bottom of surprising results. We often need to go from a vague question to a running experiment quickly.

Responsibilities:

• Design, run and interpret large-scale experiments on agent teams, reasoning rigorously about what the data does and doesn't show
• Investigate how performance and efficiency change as team size, compute and task horizon grow, and find the bottlenecks that limit them
• Build and scale the systems that run very large agent teams reliably, and debug the failures that only appear at scale
• Design evaluations for long-horizon problems, and keep their results trustworthy
• Build the tooling and metrics that let researchers see what a large agent team is doing and why
• Partner with research teams across Anthropic so they can run their own experiments on the platform, and communicate findings clearly

You may be a good fit if you:

• Have significant software engineering, ML or research engineering experience
• Have owned something substantial end to end, such as a large system, an evaluation or benchmark, an agent product, or a research project
• Genuinely enjoy both research and engineering work
• Think quantitatively about complex systems, and think twice before trusting a number
• Can work from a vague question rather than a spec
• Are results-oriented, with a bias towards flexibility and impact
• Have clear written and verbal communication
• Care about the societal impacts of your work

Strong candidates may also have:

• Experience building or operating large-scale distributed systems, such as schedulers, sandboxed code execution, or inference and RL infrastructure
• Built evaluations, benchmarks or harnesses for LLMs or agents
• Experience building complex agentic systems that use LLMs
• Experience with scaling laws or other large-scale empirical research
• A background in operations research, statistics, economics, physics, quantitative finance, or another field that models and optimizes complex systems

Strong candidates need not have:

• Formal certifications or education credentials
• Academic research experience or publication history
• Prior experience with multi-agent systems or reinforcement learning

Representative projects:

• Measure how performance scales with the number of agents on a hard problem, and explain where the curve bends and why
• Prepare our largest-ever agent run: find what breaks as team size and task length grow together, and fix it before launch
• Work out how to allocate a fixed compute budget across a team of agents to solve a problem fastest
• Build tooling that turns the activity of a large agent team into something a researcher can read in minutes
• Design a novel eval that distinguishes real gains in teamwork from artifacts of the evaluation setup

The annual compensation range for this role is listed below.

For sales roles, the range provided is the ro