ContextQMD
Back to Spring Ai Alibaba

Subagent Pattern - Tech Due Diligence Assistant

examples/multiagent-patterns/subagent/README.md

1.1.2.25.7 KB
Original Source

Subagent Pattern - Tech Due Diligence Assistant

A multi-agent example demonstrating the TaskTool pattern: a main orchestrator agent that delegates complex work to specialized sub-agents.

Overview

The Tech Due Diligence Assistant helps evaluate software projects by combining:

  • Codebase analysis: Structure, dependencies, patterns, technical debt
  • Web research: Documentation, alternatives, benchmarks, ecosystem

The main agent uses write_todos for planning and delegates to sub-agents via the Task and TaskOutput tools.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                  Tech Due Diligence Assistant                    │
│  (Orchestrator: write_todos, Task, TaskOutput, glob, grep, web)  │
└────────────────────────────┬────────────────────────────────────┘
                              │ delegates via Task tool
    ┌─────────────────────────┼─────────────────────────┬──────────────────┐
    ▼                         ▼                         ▼                  ▼
┌──────────────┐  ┌─────────────────┐  ┌─────────────────┐  ┌─────────────────────┐
│codebase-     │  │ web-researcher  │  │ general-purpose │  │ dependency-analyzer  │
│explorer      │  │ web_fetch       │  │ glob, grep, web │  │ (API-defined)        │
│glob, grep    │  │                 │  │                 │  │ glob, grep           │
│(Markdown)    │  │ (Markdown)      │  │ (Markdown)      │  └─────────────────────┘
└──────────────┘  └─────────────────┘  └─────────────────┘

Sub-Agents

Sub-agents can be defined in two ways:

1. Markdown (file-based)

AgentToolsUse Case
codebase-explorerglob_search, grep_searchFind files, search code, analyze structure
web-researcherweb_fetchFetch URLs, research docs, compare technologies
general-purposeglob_search, grep_search, web_fetchCombined code + web analysis

Defined in src/main/resources/agents/*.md with YAML front matter.

2. API (programmatic)

AgentToolsUse Case
dependency-analyzerglob_search, grep_searchAnalyze dependencies, version conflicts, outdated libs

Defined in Java via ReactAgent.builder() and registered with TaskToolsBuilder.subAgent():

java
@Bean("dependencyAnalyzerAgent")
public ReactAgent dependencyAnalyzerAgent(ChatModel chatModel, List<ToolCallback> defaultTools) {
    return ReactAgent.builder()
        .name("dependency-analyzer")
        .description("Analyzes project dependencies...")
        .model(chatModel)
        .systemPrompt(DEPENDENCY_ANALYZER_SYSTEM_PROMPT)
        .tools(globSearch, grepSearch)
        .build();
}

// In taskTools():
TaskToolsBuilder.builder()
    .subAgent("dependency-analyzer", dependencyAnalyzerAgent)  // API-defined
    .addAgentResource(resource)  // Markdown-defined
    .build();

Running

Prerequisites

  • JDK 17+
  • AI_DASHSCOPE_API_KEY environment variable set

Interactive Mode

bash
# From project root - run with interactive chat
cd /path/to/spring-ai-alibaba
AI_DASHSCOPE_API_KEY=your_key ./mvnw -f examples/multiagent-patterns/subagent/pom.xml spring-boot:run \
  -Dspring-boot.run.arguments="--subagent.run-interactive=true"

Or set in application.yml:

yaml
subagent:
  run-interactive: true

Example Prompts

  • Simple: "Find all Java files in this project"
  • Codebase: "What frameworks and dependencies does this project use?"
  • Web: "Fetch https://spring.io/projects/spring-ai and summarize its features"
  • Dependency (API sub-agent): "Analyze this project's dependencies for version conflicts and outdated libraries"
  • Combined: "Analyze this codebase for Spring usage, then research Spring AI alternatives and compare with our current setup"

Programmatic Usage

java
@Autowired
ReactAgent orchestratorAgent;

AssistantMessage response = orchestratorAgent.call(new UserMessage(
    "Analyze this codebase for technical debt and research Spring AI documentation"));

Configuration

PropertyDefaultDescription
subagent.workspace-path${user.dir}Root path for glob_search and grep_search
subagent.run-interactivefalseRun interactive chat on startup

Key Components

  • TaskToolsBuilder: Builds Task + TaskOutput tools. Supports both:
    • Markdown: addAgentResource() / addAgentDirectory() loads specs from .md files
    • API: subAgent(type, ReactAgent) registers programmatically defined ReactAgents
  • TodoListInterceptor: Injects write_todos tool and system prompt for task planning
  • Agent specs (Markdown): name, description, tools (comma-separated) in YAML front matter