Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents

Chenglin Yu, Yuchen Wang, Songmiao Wang, Hongxia Yang, Ming Li

TL;DR

InfiAgent addresses the brittleness of long-horizon LLM agents by decoupling long-term task memory from the bounded reasoning context. It externalizes persistent state into a file-centric workspace and reconstructs a fixed-size reasoning context at each step, ensuring the context length remains $\mathcal{O}(1)$ regardless of task duration. The framework employs a hierarchical, DAG-based agent architecture and an external attention pipeline to manage large information sources without inflating internal reasoning load. Experimental results on the DeepResearch benchmark and an 80-paper literature review task show that a 20B open-source model can rival larger proprietary systems and maintain high long-horizon coverage, supporting explicit state externalization as a practical foundation for stable autonomous research agents. The work demonstrates that persistent state management, when combined with structured task decomposition and external attention, enables scalable,Explainable, and robust long-horizon AI agents suitable for broad knowledge-work applications.

Abstract

LLM agents can reason and use tools, but they often break down on long-horizon tasks due to unbounded context growth and accumulated errors. Common remedies such as context compression or retrieval-augmented prompting introduce trade-offs between information fidelity and reasoning stability. We present InfiAgent, a general-purpose framework that keeps the agent's reasoning context strictly bounded regardless of task duration by externalizing persistent state into a file-centric state abstraction. At each step, the agent reconstructs context from a workspace state snapshot plus a fixed window of recent actions. Experiments on DeepResearch and an 80-paper literature review task show that, without task-specific fine-tuning, InfiAgent with a 20B open-source model is competitive with larger proprietary systems and maintains substantially higher long-horizon coverage than context-centric baselines. These results support explicit state externalization as a practical foundation for stable long-horizon agents. Github Repo:https://github.com/ChenglinPoly/infiAgent

InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents

TL;DR

InfiAgent addresses the brittleness of long-horizon LLM agents by decoupling long-term task memory from the bounded reasoning context. It externalizes persistent state into a file-centric workspace and reconstructs a fixed-size reasoning context at each step, ensuring the context length remains regardless of task duration. The framework employs a hierarchical, DAG-based agent architecture and an external attention pipeline to manage large information sources without inflating internal reasoning load. Experimental results on the DeepResearch benchmark and an 80-paper literature review task show that a 20B open-source model can rival larger proprietary systems and maintain high long-horizon coverage, supporting explicit state externalization as a practical foundation for stable autonomous research agents. The work demonstrates that persistent state management, when combined with structured task decomposition and external attention, enables scalable,Explainable, and robust long-horizon AI agents suitable for broad knowledge-work applications.

Abstract

LLM agents can reason and use tools, but they often break down on long-horizon tasks due to unbounded context growth and accumulated errors. Common remedies such as context compression or retrieval-augmented prompting introduce trade-offs between information fidelity and reasoning stability. We present InfiAgent, a general-purpose framework that keeps the agent's reasoning context strictly bounded regardless of task duration by externalizing persistent state into a file-centric state abstraction. At each step, the agent reconstructs context from a workspace state snapshot plus a fixed window of recent actions. Experiments on DeepResearch and an 80-paper literature review task show that, without task-specific fine-tuning, InfiAgent with a 20B open-source model is competitive with larger proprietary systems and maintains substantially higher long-horizon coverage than context-centric baselines. These results support explicit state externalization as a practical foundation for stable long-horizon agents. Github Repo:https://github.com/ChenglinPoly/infiAgent
Paper Structure (38 sections, 5 equations, 3 figures, 2 tables)

This paper contains 38 sections, 5 equations, 3 figures, 2 tables.

Table of Contents

  1. Keywords:
  2. Introduction
  3. Related Work
  4. Multi-Agent System Architectures
  5. Task Decomposition in Multi-Agent Systems
  6. Long-Context and Autonomous Research Agents
  7. Formalizing File-Centric State for Long-Horizon Agents
  8. Agent Execution as a State-Conditioned Decision Process
  9. Persistent State Externalization
  10. Bounded Reasoning Context Reconstruction
  11. Comparison and Implications
  12. Methodology: The InfiAgent Framework
  13. File-Centric State Management and Task Memory
  14. Periodic State Consolidation.
  15. Multi-Level Agent Hierarchy
  16. ...and 23 more sections

Figures (3)

  • Figure 1: The InfiAgent Framework. InfiAgent implements a hierarchical execution stack over a file-centric persistent state. Files serve as the authoritative task memory, while an external attention mechanism processes heavy documents outside the bounded reasoning context. Periodic state consolidation refreshes the agent’s context from the workspace snapshot.
  • Figure 2: Component-wise comparison on DeepResearch. Scores are broken down by evaluation dimension. InfiAgent shows strong performance on instruction following and readability, which are closely related to structured state management and output control.
  • Figure 3: Performance vs. model size on DeepResearch. InfiAgent (20B) achieves competitive performance relative to larger proprietary agents evaluated on the same benchmark, suggesting an improved efficiency–performance trade-off.