MCPDisCo is an ecosystem-scale reinforcement learning framework for MCP-native agents that mitigates the seesaw effect, a joint-learning failure mode, through token-masked reward attribution and bidirectional gradient surgery.