TodoEvolve: Learning to Architect Agent Planning Systems

Planning has become a central capability for contemporary agent systems in navigating complex, long-horizon tasks, yet existing approaches predominantly rely on fixed, hand-crafted planning structures that lack the flexibility to adapt to the structural diversity of open-ended problems. To address this limitation, we introduce TodoEvolve, a meta-planning paradigm that autonomously synthesizes and dynamically revises task-specific planning architectures. Specifically, we first construct PlanFactory, a modular design space that standardizes diverse planning paradigms within a unified codebase encompassing topology, initialization, adaptation, and navigation, thereby providing a common interface for heterogeneous planning patterns. Leveraging PlanFactory, we collect high-quality planning trajectories and train Todo-14B via Impedance-Guided Preference Optimization (IGPO), a multi-objective reinforcement learning objective that encourages the generation of planning systems that are performant, stable, and token-efficient across arbitrary tasks and agent backbones. Empirical evaluations on five agentic benchmarks demonstrate that TodoEvolve consistently surpasses carefully engineered planning modules while maintaining economical API costs and runtime overhead. Our codes are available at \url{https://github.com/EcthelionLiu/TodoEvolve}.

Today’s AI agents often need to solve complex, multi-step tasks, such as searching the web, using tools, writing code, or coordinating several subgoals. Most existing agents rely on fixed planning recipes designed by humans, but different tasks may require very different ways of organizing work. A plan that works well for one problem can be inefficient or unreliable for another. We introduce TodoEvolve, a system that learns to design a planning strategy for each task. Instead of using a single hand-written planner, TodoEvolve builds a task-specific planning structure and revises it as execution progresses. To make this possible, we organize existing planning methods into reusable building blocks and train a meta-planner to combine them using feedback from successful and efficient executions. Our results show that agents equipped with TodoEvolve solve more tasks across diverse benchmarks while keeping cost and latency manageable. This suggests that future AI agents can improve not only through stronger language models, but also by learning better ways to organize their own reasoning and actions.