🚀NEW COURSEVibe Coding AI Apps with Claude Code 🤖✨Enroll now
← All papers  /  Sep 20, 2026
Agents · Reinforcement Learning

HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning

First page
HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning
The curator’s take

Hongliang Wei and colleagues at Harbin Institute of Technology and Alibaba Cloud train one policy across several agent harnesses and introduce HarnessBandit, an online scheduler that picks which harness to train on at each optimizer step.

Ask this paper

Key points
01

Scheduling problem: Harnesses differ in system prompts, tool schemas, control loops and trajectory formats, and each step should use a harness that gives a useful learning signal and an update that helps the other harnesses.

02

Two signals: After each GRPO update, learnability is the mean absolute advantage on the batch and transferability is the cosine between a low-dimensional gradient sketch of the current harness and moving averages of the others.

03

Results: Training Qwen3.5-2B across six harnesses on ClawGym, HarnessBandit beats mixed-batch multi-harness training on PinchBench and on ClawEval, which holds out both tasks and a harness.

Abstract

Language-model agents are increasingly deployed through diverse harnesses that differ in system prompts, tool schemas, control loops, and trajectory formats. The same model can perform unevenly across these interfaces, making robustness to harness variation an important objective. A natural approach is to train a shared policy through multiple harnesses, but doing so introduces a scheduling problem: each training step should favor a harness that currently provides a useful learning signal while also producing an update that benefits the other harnesses. We develop HarnessBandit, an online scheduler that selects one harness per optimizer step. After a group-relative policy optimization (GRPO) update, it observes learnability -- the mean absolute advantage on the batch -- and transferability -- the cosine between a low-dimensional gradient sketch of the current harness and exponential moving averages of the remaining harnesses. The two signals are fused after pooled sliding-window min-max normalization and sampled with a visit-dependent bonus and an explicit exploration floor. We train Qwen3.5-2B across six harnesses on ClawGym and evaluate on PinchBench (held-out tasks, in-distribution OpenClaw) and ClawEval (held-out tasks and harness). HarnessBandit improves over mixed-batch multi-harness training on both benchmarks, while training diagnostics indicate that learnability and transferability provide distinct, evolving signals.

Every Monday
Get next week’s papers.
Subscribe on Substack