MagicAgent: Towards Generalized Agent Planning

TL;DR

MagicAgent introduces a scalable synthetic data framework and a two-stage training process to develop generalized agent planning models, significantly improving performance across diverse planning tasks and benchmarks.

Contribution

The paper presents MagicAgent, a new foundation model for generalized agent planning, with a synthetic data generation method and a two-stage training paradigm to enhance task generalization.

Findings

Achieves 75.1% on Worfbench and 86.9% on BFCL-v3 benchmarks.

Outperforms existing models under 100B parameters and surpasses ultra-scale models like GPT-5.2.

Demonstrates strong performance on in-house MagicEval benchmarks.

Abstract

The evolution of Large Language Models (LLMs) from passive text processors to autonomous agents has established planning as a core component of modern intelligence. However, achieving generalized planning remains elusive, not only by the scarcity of high-quality interaction data but also by inherent conflicts across heterogeneous planning tasks. These challenges result in models that excel at isolated tasks yet struggle to generalize, while existing multi-task training attempts suffer from gradient interference. In this paper, we present \textbf{MagicAgent}, a series of foundation models specifically designed for generalized agent planning. We introduce a lightweight and scalable synthetic data framework that generates high-quality trajectories across diverse planning tasks, including hierarchical task decomposition, tool-augmented planning, multi-constraint scheduling, procedural logic orchestration, and long-horizon tool execution. To mitigate training conflicts, we propose a two-stage training paradigm comprising supervised fine-tuning followed by multi-objective reinforcement learning over both static datasets and dynamic environments. Empirical results show that MagicAgent-32B and MagicAgent-30B-A3B achieve superior performance across diverse open-source benchmarks (\emph{e.g.}, $75.1\%$ on Worfbench and $86.9\%$ on BFCL-v3), as well as strong results on our in-house MagicEval benchmarks, substantially outperforming existing sub-100B models and surpassing leading ultra-scale models, including GPT-5.2, Kimi-K2 and GLM-4.7.