We are releasing Pharos, a reasoning system designed for long-horizon research on challenging, open-ended mathematical problems for which no reasonably workable route is known in advance. It inherits selected design principles from Danus and Rethlas, especially the fact graph and the multi-agent architecture, while substantially redesigning and extending the remaining system components.
Helper agents assess workers’ progress, collect and digest the literature, and scout candidate routes before a worker is committed to one. Their reports inform the main agent without entering a proof. This lets the main agent concentrate on mathematics and overall strategy, maintained on a route board that records the state of every route. It begins by surveying techniques already applied to the problem, explores known approaches, monitors progress across multiple routes, judges when a route has stalled and should be stopped, and proposes new routes from what it observes.
Workers reason more deeply, with facts representing substantial, self-contained advances. Pharos also schedules computing resources for code and local numerical computations. Its writing system produces shorter, clearer manuscripts, while a reporter agent and new skills help people follow a run and guide the system. Command-line tools coordinate its agents, and context passed to models during tool use is compressed to reduce cost and improve throughput.
Pharos is a preview and remains under active development. The GitHub repository contains the code, prompts and skills, along with setup and usage instructions.