AgentJet: A Flexible Swarm Training Framework for Agentic Reinforcement Learning

Forum topic · 小凯 · 2026-06-05

Summary

AgentJet is a distributed swarm training framework for large language model (LLM) agent reinforcement learning, proposed by Qingxu Fu, Boyin Liu, and Shuchang Tao (arXiv:2606.04484). Unlike centralized frameworks that tightly couple agent rollouts with model optimization, AgentJet uses a decoupled multi-node architecture: swarm server nodes host trainable models and run optimization on GPU clusters, while swarm client nodes execute arbitrary agents on arbitrary devices. This design enables heterogeneous multi-model reinforcement learning for multi-agent teams with multiple LLM brains, multi-task cocktail training with isolated agent runtimes, fault-tolerant execution against external environment failures, and real-time code iteration by swapping client nodes during training. A timeline-merging context tracking module deduplicates redundant context, delivering 1.5-10x training speedups in multi-model, multi-turn, multi-agent RL settings. AgentJet also introduces an automated research system that takes a research topic as input and autonomously runs multi-day RL research on large clusters, reproducing key exploratory workflows of RL researchers without human intervention.

论文概要

研究领域: ML 作者: Qingxu Fu, Boyin Liu, Shuchang Tao 发布时间: 2025-06-01 arXiv: 2606.04484

Introduction

We present AgentJet, a distributed swarm training framework for large language model (LLM) agent reinforcement learning. Unlike centralized frameworks that tightly couple agent rollouts with model optimization, AgentJet adopts a decoupled multi-node architecture in which swarm server nodes host trainable models and run optimization on GPU clusters, whereas swarm client nodes execute arbitrary agents on arbitrary devices.

Key Capabilities

This design provides capabilities that are difficult to support in centralized frameworks:

1. Heterogeneous multi-model reinforcement learning — enabling the training of heterogeneous multi-agent teams with multiple LLMs as brains. 2. Multi-task cocktail training — with isolated agent runtimes. 3. Fault-tolerant execution — preventing external environment failures from interrupting the training process. 4. Real-time code iteration — allowing agents to be edited during training by replacing swarm client nodes.

Performance

To support efficient RL in multi-model, multi-turn, and multi-agent scenarios, AgentJet introduces a timeline-merging context tracking module that consolidates redundant context and achieves 1.5-10x training speedup.

Automated Research System

Finally, AgentJet introduces an automated research system that takes a research topic as input and autonomously conducts long-horizon, multi-day RL research on large clusters. Leveraging the swarm architecture, the system reproduces key exploratory workflows of RL researchers without human intervention.

---

*Auto-collected on 2026-06-05*