Examples

CoreHIERARCHICAL

Competitive Market Analysis

Hierarchical multi-agent research. A program manager coordinates 4 specialists — market intelligence, research, strategy, and writing — to produce a comprehensive competitive landscape report.

Agents:Research Program ManagerMarket Intelligence AnalystResearch AnalystStrategy ConsultantExecutive Writer

Workflow

Program Manager
  ├──► Market Intelligence Analyst
  ├──► Research Analyst
  ├──► Strategy Consultant
  └──► Executive Writer ──► Report

Hierarchical delegation5 specialist agentsManager coordinationCompetitive intelligence

CorePARALLEL

code Source

Investment Due Diligence

Parallel due diligence across financial, market, and legal dimensions. Specialist agents analyze a company simultaneously, then a director synthesizes findings into a comprehensive assessment.

Agents:Program DirectorFinancial AnalystMarket AnalystLegal & Regulatory Analyst

Workflow

┌─ Financial Analyst  ─┐
├─ Market Analyst     ─┤──► Program Director ──► Report
└─ Legal & Regulatory ─┘

Parallel executionLayer synchronization4 concurrent agentsResult synthesis

CoreHIERARCHICAL

code Source

Hierarchical Web Research

Manager agent creates an execution plan, delegates search tasks to specialists, collects results, and orchestrates the final report. Workers are selected based on task requirements and tool availability.

Agents:Research ManagerWeb SearcherData AnalystFact CheckerReport Writer

Workflow

Manager ──► Plan
  ├──► Web Searcher ──► raw data
  ├──► Data Analyst ──► insights
  ├──► Fact Checker ──► verified
  └──► Report Writer ──► final report

Manager delegationDynamic task assignmentAgent selection by capability5 agents

AdvancedSELF_IMPROVING

code Source

Dynamic Skill Generation

The analyst executes tasks while the reviewer evaluates output quality. When capability gaps are detected, new Groovy-based skills are generated, validated in a sandbox, and hot-loaded into the agent mid-run. RL policy decides when to converge.

Agents:AnalystReviewer

Workflow

Analyst ──► Output ──► Reviewer
  ▲                        │
  │   ┌─ CAPABILITY_GAP ◄─┘
  │   ▼
  │  Generate Skill ──► Validate ──► Register
  └──────────── Retry with new skill

Dynamic skill generationGroovy sandbox validationRL convergence policySkill persistence

AdvancedSELF_IMPROVING

code Source

Self-Evolving Swarm

Transparent self-evolution: the swarm discovers a better architecture and applies it automatically on the next run. Configured as SEQUENTIAL, Swarm.kickoff() reads evolution history from H2 and transparently switches to PARALLEL — zero code changes needed.

Agents:Technology AnalystMarket AnalystRisk Analyst

Workflow

Run 1: [Tech] ──► [Market] ──► [Risk]  (SEQUENTIAL)
  └── observes: PROCESS_SUITABILITY
  └── persists: PROCESS_TYPE_CHANGE → H2

Run 2: Swarm.kickoff() reads H2
  ┌─ [Tech]   ─┐
  ├─ [Market] ─┤  (PARALLEL — evolved!)
  └─ [Risk]   ─┘

Transparent evolutionH2 persistent learningProcess type optimizationEvolutionAdvisor on kickoff()

EnterpriseSEQUENTIAL

code Source Enterprise

Governed Enterprise Workflow

Production-grade workflow with multi-tenancy (tenant-scoped memory/quotas), budget tracking ($5 cap, 500K token limit), and human-in-the-loop approval gates between research and writing phases.

Agents:Research AnalystReport Writer

Workflow

Tenant: acme-research | Budget: $5.00 / 500K tokens

Research Analyst ──► [APPROVAL GATE] ──► Report Writer
                     │ auto-approve    │
                     │ after 5s        │
                     └─────────────────┘
──► Budget Snapshot logged

Multi-tenancyBudget HARD_STOP/WARNApproval gatesTenant quotas

AdvancedSWARM

code Source

Competitive Research Swarm

Distributed fan-out pattern. Discovery phase identifies competitor targets, then parallel agents research each target independently. A coordinator synthesizes all findings into a master report.

Agents:Discovery AgentResearch AgentSynthesis Agent

Workflow

Discovery ──► [Target A, Target B, Target C]
  ├──► Research Agent (A) ──┐
  ├──► Research Agent (B) ──┤──► Coordinator ──► Master Report
  └──► Research Agent (C) ──┘

Distributed fan-outParallel sub-swarmsShared skill registryMaster synthesis

CoreITERATIVE

code Source

Iterative Investment Memo

Execute-review-refine loop. The analyst researches, the memo author drafts, and the managing director reviews against a 7-point rubric. Iterates until approved or max iterations reached.

Agents:Equity Research AnalystMemo AuthorManaging Director

Workflow

Research Analyst ──► Memo Author ──► Managing Director
  ▲                                        │
  │         NEEDS_REFINEMENT ◄─────────────┘  (with feedback)
  │                │
  └────────────────┘

  APPROVED ──► Final Memo

Quality-driven iteration7-point review rubricMax iteration cap3-agent collaboration

DSLALL

code Source

YAML DSL Workflows

30+ YAML workflow definitions covering every process type, budget tracking, governance gates, tool hooks, conditional tasks, graph workflows with state channels, and composite pipelines. Zero Java code required.

Agents:Defined in YAML

Workflow

YAML File ──► YamlSwarmParser ──► SwarmDefinition
                                      │
                                      ▼
                               SwarmCompiler ──► Live Swarm
                                      │
                                      ▼
                                swarm.kickoff()

Template variables {{topic}}All 7 process typesTool hooks in YAMLGraph conditional routing

AdvancedGRAPH

code Source

Human-in-the-Loop Graph

Graph workflow with conditional routing. Writer produces draft, human reviewer scores it. If score >= 80, proceed to publish. If iteration >= 3, force publish. Otherwise, loop back to editor for refinement.

Agents:WriterHuman ReviewerEditor

Workflow

START ──► Writer ──► Reviewer
                       │
            score >= 80 ├──► END (publish)
         iteration >= 3 ├──► END (force)
              default   └──► Editor ──► Writer (loop)

Conditional edgesState channels (score, iteration)Human review checkpointLoop with exit conditions

CoreSEQUENTIAL

code Source

Data Pipeline

ETL-style pipeline. Collector ingests data from CSV and database tools, analyst performs statistical analysis, and generator produces a formatted report with charts description.

Agents:Data CollectorData AnalystReport Generator

Workflow

Data Collector ──[csv, database]──► Data Analyst ──[calculator]──► Report Generator ──► output/

CSV analysis toolDatabase query toolCalculator toolFile output

CoreGRAPH

code Source

Customer Support REST API

Full REST API application with AI-powered chat, intelligent routing via SwarmGraph, conversation history, product catalog, order management, and ticket system. Runs on port 8080 with a web UI.

Agents:ClassifierBilling AgentTechnical AgentAccount Agent

Workflow

REST API :8080 ──► Classifier
  ├── BILLING ──► Billing Agent ──┐
  ├── TECHNICAL ──► Tech Agent   ──┤──► QA ──► Response
  ├── ACCOUNT ──► Account Agent ──┤
  └── GENERAL ──► General Agent ──┘

REST APISwarmGraph routingConversation historyWeb UI

CoreSEQUENTIAL

code Source

RAG Knowledge Base

Complete RAG application with document ingestion, vector store integration (Chroma), semantic search, and multi-agent Q&A pipeline. Runs as a REST API on port 8080.

Agents:Retriever AgentWriter Agent

Workflow

POST /ingest ──► Embed & Store
POST /ask ──► Retriever ──► (Vector Store) ──► Writer ──► Response

Document ingestionVector store (Chroma)Semantic searchREST API

EnterpriseSEQUENTIAL

code Source

SecureOps Assessment

Security assessment pipeline with budget controls and tool permission enforcement. Recon agent uses web scraping (READ_ONLY), vulnerability analyst processes findings, writer generates the assessment report.

Agents:Recon AgentVulnerability AnalystReport Writer

Workflow

Recon Agent ──[web-scrape: READ_ONLY]──► Vuln Analyst ──► Report Writer

Budget: $5 / 200K tokens | Permission: READ_ONLY enforced

Tool permission levelsBudget trackingREAD_ONLY enforcementSecurity-focused

EnterpriseSWARM

code Source

Distributed Penetration Test

Swarm-based security testing. Scanner discovers targets, parallel agents assess each target independently, coordinator produces a comprehensive security report.

Agents:ScannerExploit AnalystReport Writer

Workflow

Scanner ──► [Target 1, Target 2, ...]
  ├──► Exploit Agent ──┐
  ├──► Exploit Agent ──┤──► Security Report
  └──► Exploit Agent ──┘

SWARM processParallel security agentsCoverage enforcementComprehensive report

Getting StartedSEQUENTIAL

code Source

Hello World — Single Agent

The simplest possible SwarmAI setup: one agent, one task, sequential process, no tools. Start here to learn the basics of Agent, Task, and Swarm.

Agents:Summarizer

Workflow

[Summarizer] ──► output

1 agent1 taskNo toolsMinimal setup

Getting StartedSEQUENTIAL

code Source

Agent with Tool Calling

A single agent equipped with the CalculatorTool to perform precise arithmetic. The agent decides when to call the tool based on the task description.

Agents:Math Tutor

Workflow

[Math Tutor] ──uses──► (CalculatorTool) ──► output

Tool registrationSpring-managed toolsTool hooksAuto tool invocation

Getting StartedSEQUENTIAL

code Source

Agent-to-Agent Task Handoff

Two agents in sequence. The researcher gathers information, then the editor refines it. Demonstrates task dependencies where one agent output feeds into the next.

Agents:ResearcherEditor

Workflow

[Researcher] ──► [Editor] ──► output

Task dependenciesmaxTurns controlPermission levelsOutput chaining

Getting StartedSEQUENTIAL

code Source

Shared Context Between Agents

Three agents in a pipeline sharing context (topic, audience, tone, word count) through the inputs map. Each agent builds on the previous output.

Agents:OutlinerDrafterPolisher

Workflow

[Outliner] ──► [Drafter] ──► [Polisher] ──► output
       └──── shared context ────┘

Inputs mapContext variables3-stage pipelinedependsOn chaining

Getting StartedSEQUENTIAL

code Source

Multi-Turn Deep Reasoning

A single agent that reasons across multiple LLM turns with automatic context compaction. The agent autonomously decides when to continue and when analysis is complete.

Agents:Deep Researcher

Workflow

[Deep Researcher]
  turn 1 → turn 2 → ... → turn 5
  (auto-compact after 4K tokens)
  ──► output

maxTurns(5)CompactionConfigCONTINUE/DONE markersIterative reasoning

CoreSEQUENTIAL

code Source

Streaming Real-Time Responses

Reactive multi-turn execution with progress hooks showing incremental output as the agent reasons through the problem.

Agents:Streaming Agent

Workflow

[Agent] ──► stream chunk 1 ──► chunk 2 ──► ... ──► final output

Reactive streamingProgress hooksMulti-turnIncremental output

CoreSEQUENTIAL

code Source

Error Handling & Recovery

3 resilience scenarios: tool failure recovery, budget enforcement (HARD_STOP), and timeout handling. Demonstrates how SwarmAI handles failures gracefully.

Agents:Various

Workflow

Scenario 1: Tool failure ──► recovery
Scenario 2: Budget exceeded ──► HARD_STOP
Scenario 3: Timeout ──► graceful shutdown

Tool failure recoveryBudget HARD_STOPTimeout handlingGraceful degradation

CoreSEQUENTIAL

code Source

Conversation Memory Persistence

Shared InMemoryMemory across agents — save, search, recall, and cross-agent knowledge sharing. Demonstrates persistent context across the workflow.

Agents:Research AgentAnalysis Agent

Workflow

[Research Agent] ──save──► (Memory) ──recall──► [Analysis Agent]

InMemoryMemorySave & recallCross-agent sharingKnowledge persistence

CoreSEQUENTIAL

code Source

Multi-LLM Provider Switching

Same task executed at different temperatures and model variants, with side-by-side comparison of outputs. Useful for prompt engineering and model evaluation.

Agents:Agent (multiple configs)

Workflow

[Task] ──► Agent (temp=0.1) ──► output A
       ──► Agent (temp=0.7) ──► output B
       ──► Compare results

Temperature comparisonModel variantsSide-by-side outputPrompt engineering

CorePARALLEL

code Source

Multi-Language Translation

3 agents analyze the same topic in English, Spanish, and French simultaneously. A synthesizer produces a cross-cultural report combining all perspectives.

Agents:English AgentSpanish AgentFrench AgentSynthesizer

Workflow

┌─ English Agent  ─┐
├─ Spanish Agent  ─┤──► Synthesizer ──► Cross-Cultural Report
└─ French Agent   ─┘

Parallel execution3 languagesCross-cultural synthesisMulti-agent collaboration

CorePARALLEL

code Source

Stock Market Analysis

3 analyst agents research a stock in parallel using web search and SEC filings tools, then an investment advisor synthesizes findings into a recommendation.

Agents:Financial AnalystResearch AnalystFilings AnalystInvestment Advisor

Workflow

┌─ Financial Analyst ──[calculator, web]──┐
├─ Research Analyst ──[web, sec-filings]──┤──► Investment Advisor ──► Report
└─ Filings Analyst ──[web, sec-filings]──┘

Parallel analysisSEC filings toolWeb searchInvestment recommendation

CoreSEQUENTIAL

code Source

Codebase Analysis

Analyze codebase architecture, metrics, and dependencies. Works on any local code directory without external API keys.

Agents:Code Analyst

Workflow

[Code Analyst] ──[file-read, directory-read]──► Architecture Report

Local analysisNo API keys neededFile toolsArchitecture insights

CoreSEQUENTIAL

code Source

MCP Model Context Protocol

Research workflow using MCP (Model Context Protocol) tools for web fetch and search. Demonstrates the MCP integration pattern.

Agents:Research Agent

Workflow

[Research Agent] ──[mcp-web-fetch, mcp-web-search]──► Report

MCP integrationWeb fetch toolWeb search toolProtocol-based tools

CoreSEQUENTIAL

code Source

Scheduled Cron Monitoring

3-iteration monitoring with file-based state. Detects trends across runs using persistent state files.

Agents:Monitor Agent

Workflow

[Monitor] ──► Run 1 ──► Run 2 ──► Run 3
                └── file state ──┘ (trend detection)

Scheduled executionFile-based stateTrend detectionMulti-iteration

CoreGRAPH

code Source

Workflow Visualization

Build 4 graph topologies and generate Mermaid diagrams. No LLM needed — demonstrates the graph API and visualization capabilities.

Agents:None (diagram generation)

Workflow

SwarmGraph.create() ──► topology ──► Mermaid diagram

Mermaid diagrams4 graph topologiesNo LLM neededVisual workflow design

AdvancedGRAPH

code Source

Evaluator-Optimizer Feedback Loop

Generate → evaluate → optimize loop with a quality gate. The evaluator scores output on multiple criteria; if score >= 80, it passes. Otherwise, the optimizer refines and loops back.

Agents:GeneratorEvaluatorOptimizer

Workflow

Generator ──► Evaluator (score >= 80?)
  ▲                │
  │   NO ◄──────────┘
  │   ▼
  └── Optimizer

  YES ──► Final Output

Quality gateScore thresholdOptimize loopMulti-criteria evaluation

AdvancedGRAPH

code Source

Multi-Agent Debate

Two agents debate a proposition over 3 rounds, then a judge declares the winner. Demonstrates the peer interaction pattern with structured argumentation.

Agents:ProponentOpponentJudge

Workflow

Proponent ──► Opponent ──► Proponent ──► ... (3 rounds)
                                                    ▼
                                               [Judge] ──► Winner

Debate roundsPeer interactionJudge evaluationStructured argumentation

AdvancedSEQUENTIAL

code Source

Unit Testing Agents with Mocks

Agent output quality evaluation with 5-criterion scoring. Includes JUnit 5 unit tests using mock ChatClient — no LLM needed for testing.

Agents:Test Agent

Workflow

[Agent] ──► output ──► 5-criterion scorer

JUnit 5: mock ChatClient ──► verify agent config, hooks, dependencies

Mock ChatClientJUnit 5 tests5-criterion scoringNo LLM needed for tests

AdvancedSWARM

code Source

Competitive Research Swarm

Distributed fan-out pattern. Discovery phase identifies competitor targets, then parallel agents research each independently. Coordinator synthesizes into a master report.

Agents:Discovery AgentResearch AgentSynthesis Agent

Workflow

Discovery ──► [Target A, Target B, Target C]
  ├──► Research Agent (A) ──┐
  ├──► Research Agent (B) ──┤──► Coordinator ──► Master Report
  └──► Research Agent (C) ──┘

Distributed fan-outParallel sub-swarmsShared skill registryMaster synthesis

AdvancedSWARM

code Source

Investment Analysis Swarm

Multi-company investment analysis with parallel agents and cross-agent skill sharing. Each company is analyzed independently, then findings are synthesized.

Agents:Discovery AgentAnalysis AgentCoordinator

Workflow

┌─ Company A Agent ──┐
├─ Company B Agent ──┤──► Coordinator ──► Investment Report
└─ Company C Agent ──┘

Multi-company analysisParallel agentsCross-agent skillsInvestment synthesis

EnterpriseSEQUENTIAL

code Source Enterprise

Enterprise Governance & SPI Hooks

Enterprise-grade workflow with SPI extension points (AuditSink, LicenseProvider, MeteringSink), multi-tenancy isolation, and human-in-the-loop approval gates.

Agents:Research AnalystReport Writer

Workflow

[Researcher] ──► [Approval Gate] ──► [Writer]
     │                                    │
     +── AuditSink ─── MeteringSink ──────+
     +── TenantContext ── BudgetTracker ──+

SPI extension pointsAuditSinkLicenseProviderMeteringSinkMulti-tenancy

EnterpriseSEQUENTIAL

code Source

Audit Trail Research Pipeline

Research pipeline with full observability. Every tool call is audited and sanitized, the entire workflow is recorded for replay. Multi-turn reasoning with auto-compaction.

Agents:Research Agent

Workflow

[Researcher] ──[audit + sanitize + rate-limit]──► output
     └── Decision Tracing ── Event Replay ── Structured Logging

Audit hooksSanitizationRate limitingDecision tracingEvent replay

EnterpriseCOMPOSITE

code Source

Governed Pipeline with Checkpoints

Multi-stage composite pipeline: Parallel research → Hierarchical synthesis → Iterative review. Checkpoints between stages, budget enforcement, and Mermaid diagram generation.

Agents:Multiple (3-stage)

Workflow

[Parallel Research] ──► checkpoint ──► [Hierarchical Synthesis]
     ──► checkpoint ──► [Iterative Review] ──► Final Report

Composite processCheckpointsBudget enforcementMermaid diagrams3-stage pipeline

CoreSEQUENTIAL

code Source

RAG Retrieval-Augmented Research

RAG workflow with vector store search and multi-agent evidence-grounded report writing using InMemoryKnowledge and SemanticSearchTool.

Agents:Retriever AgentWriter Agent

Workflow

[Retriever] ──(Vector Store)──► [Writer] ──► Grounded Report

InMemoryKnowledgeSemanticSearchToolEvidence-groundedVector store search

AdvancedSELF_IMPROVING

code Source

Deep Reinforcement Learning (DQN)

Iterative workflow powered by a Deep Q-Network (DQN) policy engine. The RL agent learns optimal strategies for task execution through experience replay and neural network optimization.

Agents:AnalystReviewer

Workflow

Analyst ──► Reviewer ──► DQN Policy
  ▲                         │
  └── optimize strategy ────┘

DQN policy engineExperience replayNeural networkReinforcement learning

Example Workflows

Competitive Market Analysis

Investment Due Diligence

Hierarchical Web Research

Dynamic Skill Generation

Self-Evolving Swarm

Governed Enterprise Workflow

Competitive Research Swarm

Iterative Investment Memo

YAML DSL Workflows

Human-in-the-Loop Graph

Data Pipeline

Customer Support REST API

RAG Knowledge Base

SecureOps Assessment

Distributed Penetration Test

Hello World — Single Agent

Agent with Tool Calling

Agent-to-Agent Task Handoff

Shared Context Between Agents

Multi-Turn Deep Reasoning

Streaming Real-Time Responses

Error Handling & Recovery

Conversation Memory Persistence

Multi-LLM Provider Switching

Multi-Language Translation

Stock Market Analysis

Codebase Analysis

MCP Model Context Protocol

Scheduled Cron Monitoring

Workflow Visualization

Evaluator-Optimizer Feedback Loop

Multi-Agent Debate

Unit Testing Agents with Mocks

Competitive Research Swarm

Investment Analysis Swarm

Enterprise Governance & SPI Hooks

Audit Trail Research Pipeline

Governed Pipeline with Checkpoints

RAG Retrieval-Augmented Research

Deep Reinforcement Learning (DQN)

Run an Example