Applied Research - Agentic RL

Prime Intellect is building the open superintelligence stack: the infrastructure frontier AI labs build internally, made available to every ambitious AI team.

Our platform, Lab, unifies compute, environments, evaluations, secure sandboxes, high-performance training, and deployment into one full-stack system for post-training at frontier scale - from SFT and RL to tool use, agent workflows, and continuously improving production models. We are building open frontier AI: open-source models trained end to end for long-horizon tasks like autonomous research, and the full-stack platform our own research team uses to build them. The next generation of AI companies, enterprises, and research teams do not just need more GPUs. They need the ability to turn their own workflows, tools, data, and feedback loops into superintelligence they own.

Prime Intellect has raised $150M in total funding from Founders Fund, Radical Ventures, NVIDIA, and exceptional AI, infrastructure, and enterprise operators — including Andrej Karpathy, Dwarkesh Patel, and leaders and founders from Ramp, Perplexity, Harvey, Mercor, Zapier, Datadog, Cognition, OpenAI, Thinking Machines, Together AI, SemiAnalysis, LangChain, Browserbase, Cloudflare, Sierra, Databricks, Airbnb, OpenRouter, Standard Intelligence, Fleet, Core Auto, and more. We are looking for people who want to build at the intersection of frontier research, real infrastructure, and go-to-market for a category that does not fully exist yet.

You will work on reinforcement learning for agents that reason over long horizons, use tools, coordinate with other agents, recover from failure, and improve from experience. This is Prime's broad, cross-domain Applied Research role, focused on RL and ML systems. Your work should hold up across several application domains: algorithms, environments, harnesses, reward and credit-assignment methods, evaluation, and scalable experimental systems. You will work regularly with application teams, partners, and open-source users while remaining focused on methods and systems that transfer across domains.

• →

  RL and post-training methods for tool-using, multi-agent, memory-augmented, and long-horizon systems.

• →

  Reusable environments, harnesses, rewards, and evaluations.

• →

  Experiments that connect algorithm choices to training and inference behavior.

• →

  Open-source systems and papers built from problems surfaced by application researchers and external users.

We turn a workflow into tasks, environments, evaluations, and reward signals, run post-training experiments against real baselines, and deploy what works. When a result looks good, we find out whether the gain came from the model or from the data, harness, inference, or infrastructure. The reusable parts become shared research, open-source systems, and product capabilities.

Every Applied Research role is externally facing. Most researchers work directly with customers, users, or research partners while maintaining a substantive research agenda.

Recent Prime work includes verifiers v1, an algorithms layer for prime-rl, multi-agent training, Prime Agent, and RL systems tested at trillion-parameter scale. You will extend this line of work and pressure-test it against live application programs.

• Build agentic judges and credit-assignment methods that inspect complete trajectories, use tools, and assign useful learning signal beyond one final score.

• Make multi-agent and long-horizon environments train reliably at scale while preserving debuggable traces and comparable baselines.

• Detect when apparent gains come from reward hacking, harness behavior, data leakage, or infrastructure artifacts, then redesign the loop so the result survives.

• Relevant experience: hands-on RL training, scalable ML systems, agent environments, evaluation, inference/training infrastructure, or research crossing methods and systems.

• Hands-on experience building and debugging model-training, post-training, agent, evaluation, or closely adjacent experimental loops.

• Strong experimental judgment: meaningful baselines, careful trace inspection, clear failure attribution, and a willingness to change direction when the evidence changes.

• Strong Python and the ability to build artifacts that other researchers and engineers can run, inspect, and extend.

• Comfort owning ambiguous problems and keeping partners and teammates current as you resolve them.

We value publications, open-source systems, benchmarks, and shipped products as evidence of exceptional work. Show us something difficult you built, what failed, and how the evidence changed your approach.