MagicAgent - A basic intelligent agent model jointly launched by Honor and Fudan University

What is MagicAgent?

MagicAgent is a foundational agent model developed by Honor in collaboration with Fudan University. It achieves top performance with over 10 billion parameters using a 32B dense architecture and a 30B-A3B MoE architecture. The model covers five major scenarios—task decomposition, tool planning, and multi-constraint scheduling—through a lightweight synthetic data framework. It employs a two-stage training approach of "SFT + multi-objective reinforcement learning" to resolve multi-task conflicts and innovatively proposes the χPO algorithm to balance exploration and utilization. The model surpasses models with hundreds of billions of parameters, such as GPT-5.2 and Kimi-K2, on benchmarks like Worfbench and BFCL-v3. It is the industry's first agent model to support generalized planning across all scenarios and has been deployed in the Honor Magic series of mobile phones.

Main functions of MagicAgent

• Hierarchical task decompositionThe model can break down complex user instructions into a sequence of executable subtasks, supporting multi-step dependencies and parallel execution.
• Tool Enhancement PlanningDynamically call external APIs and tools to complete tasks that require real-time information or external capabilities through a reasoning-action loop.
• Multi-constraint scheduling: Handling planning with multiple constraints such as time, space, and resources, such as itinerary planning and meeting arrangements.
• Program Logic ArrangementThe model can understand and execute workflows with conditional branches and loop structures, and maintain complex dependencies between tasks.
• Long-term tool executionMaintain state tracking throughout dozens of interactions to stably complete long-term tasks that require multiple toolchain calls.

MagicAgent's technical principles

• Synthetic data generationMagicAgent builds a dependency graph and parameter sharing graph for its tools, defines atomic plans as the smallest semantic units, and synthesizes complex trajectories through operations such as concatenation, aggregation, and grouping. This replaces high-cost sandbox simulations, ensuring rigorous data logic and coverage of diverse scenarios.
• Two-stage training paradigmThe first stage employs supervised fine-tuning based on novelty sampling to balance the data distribution across multiple tasks. The second stage introduces a unified multi-objective reward function that combines format correctness with task semantic accuracy, and gradually improves generalization ability through offline GRPO and online χPO reinforcement learning.
• χPO algorithmThe algorithm is designed with a three-layer mechanism for sparse reward environments: token-level entropy regularization promotes vocabulary diversity exploration, thought-action separation entropy smoothing allows for high uncertainty in the reasoning stage while constraining the decision-making stage, and information bottleneck compresses redundant reasoning and retains key decision-making information, thus achieving a dynamic balance between exploration and utilization.
• MoE Load BalancingIt adopts global batch statistics to replace micro-batch constraints, supporting the natural differentiation of experts in the task dimension; combined with z-loss to suppress extreme values of routing logits, it solves the problems of expert collapse and parameter idleness in multi-task training, and maintains the decoupling of inference efficiency and model capacity.

MagicAgent project address

• arXiv technical paperhttps://arxiv.org/pdf/2602.19000

Application scenarios of MagicAgent

• Intelligent device controlUsers can use natural language commands to control their phones to complete complex operations, such as "booking a hotel by West Lake on Meituan and finding highly-rated restaurants nearby." The model automatically breaks down the task and calls the corresponding App interfaces to execute it.
• Enterprise Process AutomationIn customer service scenarios, it handles cross-system business, such as automatically querying inventory status, verifying payment information, initiating a refund process, and simultaneously notifying the user based on the order number and refund reason provided by the user.
• Personalized travel planningIt automatically generates complete itinerary plans that meet multiple constraints, such as "direct flights, number of days spent in specific cities, and consecutive trips," by combining user time windows, budget constraints, and preference tags.
• Multi-agent task orchestrationIt can serve as a central node to analyze the intentions of higher-level personnel, breaking down "preparing for a product launch" into parallel sub-tasks such as venue, materials, and guests, distributing them to professional intelligent agents for execution, and integrating the results.
• Long-range interactive decision makingThe model can continuously track the state in dozens of rounds of dialogue, handle chain-dependent tasks such as "check flights first, then book hotels, and finally rent cars", and dynamically adjust subsequent plans based on intermediate results.