Harness-RL: Black-Box RL for Central-Agent Multi-Agent Harnesses

Peking University tackles training the central policy of a multi-agent harness without hand-integrating each harness into an RL framework. Two structural problems: an action label is a low-cardinality decision while its arguments are a high-dimensional sequence (a shared sequence-level signal yields conflicting gradients), and dynamic scheduling produces branching, parallel, context-rewritten sessions that don't flatten into one token sequence. Harness-RL pairs Conflict-Aware Policy Optimization (CAPO, decoupling action from args) with interface-level black-box trajectory construction captured at the harness boundary.

Sits at the harness×RL intersection alongside Microsoft's Agent Lightning (which owns the same boundary for arbitrary harnesses) and Peking's own Harness-Bench.