fathom.party

We enable models to hillclimb non-computationally verifiable domains through a co-evolutionary training setup, where a Dungeon Master agent procedurally generates increasingly difficult escape-room worlds and a Hero agent learns to solve them through long-horizon, tool-based reasoning, external scratchpad planning, and recovery from uncertainty, with reward signals that shape exploration, inference, and correct execution. We then validate that the learned behavior is not just map-specific by swapping the world surface from game objects (rooms/items/guards) to application-style entities (apps/docs/messages/apis) while keeping the same planning and interaction loop, showing measurable gains in reward and success under harder environments and richer task variants.