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

TL;DR

Self-Manager introduces a parallel agent loop for long-form research, enabling asynchronous, concurrent execution with isolated contexts, significantly improving scalability, efficiency, and performance over traditional sequential agents.

Contribution

It presents a novel parallel agent loop architecture with isolated contexts and thread management, enhancing scalability and effectiveness in complex research tasks.

Findings

Outperforms single-agent baselines across all metrics

Improves contextual capacity, efficiency, and generalization

Demonstrates the importance of parallel execution design choices

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.