In this work the authors introduce HarnessNet, a framework for hierarchical harness evolution in multi‑agent systems (MAS). The elementary component, called an Agentron, consists of an evolvable forward harness, a frozen foundation model, and a local optimizer that can only observe and modify that harness. Agentrons are linked in a layered directed graph; environment feedback propagates backward through the graph and upstream optimizers receive node‑specific feedback from the downstream layer. This decomposition turns a monolithic joint‑harness optimization into many smaller tasks with reduced context and workload, while keeping access to internal trajectories and raw feedback local, thus preserving privacy. Experiments on AgentBench with several foundation models show that HarnessNet attains performance gains comparable to a single global optimizer but uses far fewer tokens and less time per optimizer call. A migration study demonstrates that the evolved harnesses remain effective after swapping the foundation models, indicating cross‑model transferability.
Review