flash-searcher-dag-parallel-agents

name	flash-searcher-dag-parallel-agents
title	Flash-Searcher: DAG-Based Parallel Execution for LLM Web Agents
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2509.25301
keywords	["agent-orchestration","parallel-execution","DAG","web-agents","efficiency"]
description	Reduce agent execution steps by 35% and latency by parallelizing sequential tool calls through task dependency graphs (DAGs). Use when deploying information-retrieval agents where tool execution ordering is flexible.

Flash-Searcher: DAG-Based Parallel Execution for LLM Web Agents

Flash-Searcher replaces sequential agent reasoning with concurrent subtask execution via dependency-aware directed acyclic graphs (DAGs). This framework decomposes complex agent tasks into independent execution paths while maintaining logical coherence, reducing steps by 35% and improving overall latency.

Core Architecture

• DAG decomposition: Analyzes task dependencies to identify independent subtasks
• Dynamic workflow optimization: Adjusts parallelization strategy based on runtime execution
• Dependency tracking: Ensures downstream tasks receive upstream results appropriately
• Lightweight fine-tuning: Adaptable to smaller models (8B parameters)

Implementation Steps

Setup DAG-based agent execution framework:

# Initialize DAG-based parallel agent executor
from flash_searcher import DAGAgentExecutor, DependencyAnalyzer

executor = DAGAgentExecutor(
    model=your_llm,
    max_parallel_tasks=4,
    task_timeout=30,
    execution_strategy="dynamic"
)

# Create dependency analyzer for task decomposition
analyzer = DependencyAnalyzer(
    llm=your_llm,
    tool_catalog=web_tools
)

Execute task decomposition and parallel execution:

# Decompose task into DAG structure
task = "Find the top 3 restaurants in NYC with highest ratings, then check their hours"

dag = analyzer.decompose(
    task=task,
    tools=["search", "get_info", "check_hours"]
)

# DAG structure shows parallelizable stages:
# Stage 1: search(restaurant_query)  [parallel with start]
# Stage 2: get_info(restaurant_results)  [parallel with start]
# Stage 3: check_hours(hours_info)  [depends on Stage 2]

# Execute with parallelization
results = executor.execute(
    dag=dag,
    max_parallel=4,
    dynamic_optimization=True
)

# Return integrated results (67.7% success rate on BrowseComp)
()