Teams at White Circle

Design and run adversarial single- and multi-agent environments to find concrete failure modes in LLM agents. Orchestrate large-scale experiments against external APIs and internal models, instrument emergent behaviors, catalogue failures, and translate findings into internal models and public research.

Design and run empirical experiments to find how LLM agents fail in realistic settings; build automated audit agents, fine-tune and run models, perform white-box and black-box investigations, produce evals and publish findings that feed into product guardrails.

Build, own, and maintain internal benchmark suites for single- and multi-turn content and agentic guardrails; create benchmarks distinguishing model capabilities; extend evals to new verticals and product features; collaborate on research quantifying realistic LLM and agent failure modes; support product integration of core model functionality.