Self-Manager: Parallel Agent Loop for Long-form Deep Research

2026-01-25

AI Summary

Existing agents struggle with long-horizon, in-depth research tasks due to limitations imposed by a single context window and sequential execution, making it difficult to simultaneously achieve focus, scalability, and adaptability. This work proposes Self-Manager, a parallel agent loop architecture that enables context isolation and dynamic scheduling. The framework dynamically spawns child threads—each with its own independent context—from a main thread and employs Thread Control Blocks (TCBs) to manage asynchronous concurrency and iterative refinement. By breaking away from the conventional single-agent sequential paradigm, Self-Manager achieves substantial improvements in context capacity, execution efficiency, and generalization capability, significantly outperforming current baselines on the DeepResearch Bench benchmark.

Technology Category

• Multiagent Systems: Adversarial Agents
• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Search and Optimization: Learning to Search

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Agentic search
• Economics, Online Markets and Human Computation: Architectures and workflows that use LLMs for crowd work

Abstract

Long-form deep research requires multi-faceted investigations over extended horizons to get a comprehensive report. When handling such complex tasks, existing agents manage context at the subtask level to overcome linear context accumulation and information loss. However, they still adhere to a single context window and sequential execution paradigm, which results in mutual interference and blocking behavior, restricting scalability and adaptability. To address this issue, this paper introduces Self-Manager, a parallel agent loop that enables asynchronous and concurrent execution. The main thread can create multiple subthreads, each with its own isolated context, and manage them iteratively through Thread Control Blocks, allowing for more focused and flexible parallel agent execution. To assess its effectiveness, we benchmark Self-Manager on DeepResearch Bench, where it consistently outperforms existing single-agent loop baselines across all metrics. Furthermore, we conduct extensive analytical experiments to demonstrate the necessity of Self-Manager's design choices, as well as its advantages in contextual capacity, efficiency, and generalization.

Problem

• long-form deep research
• agent context management
• sequential execution
• scalability
• mutual interference

Innovation

• parallel agent loop
• asynchronous execution
• isolated context
• Thread Control Blocks
• long-form deep research