Build reliable agentic systems with Dapr Workflow

This repository demonstrates how to build reliable agentic systems by combining Dapr Workflow orchestration with LLM-powered activities using the Dapr Conversation API.

Overview

The demos showcase the patterns described in Building Effective Agents by Anthropic, implemented using Dapr Workflow to orchestrate reliable multi-step processes and the Dapr Conversation API to interact with LLMs. Ollama is used as the local LLM provider for all demos.

Benefits of Dapr Workflow and Conversation API

Dapr Workflow

• Deterministic control over task execution with structured error handling and retries
• Durable state management that persists across failures and restarts
• Built-in orchestration patterns for complex multi-step processes
• Reliable execution with automatic recovery and compensation logic

Dapr Conversation API

• Flexibility and natural language understanding through LLM integration
• Provider-agnostic interface supporting OpenAI, Azure OpenAI, Ollama, and more
• Simplified LLM integration with consistent API across different providers
• Built-in caching to optimize performance and reduce costs

Combined Power

By combining deterministic workflows with the probabilistic nature of LLMs, you can build agentic systems that are both reliable and flexible, leveraging the best of both worlds.

Agentic Patterns

Pattern	Demo Project
Prompt Chaining	AnomalyAnalysis
Routing	GalacticAnomalyClassifier
Parallelization	StarshipDiagnostics
Orchestrator-Workers	SpaceColonyPlanner
Evaluator-Optimizer	AlienTranslator
Autonomous Agent	SpaceDebrisAgent

Pattern Architectures

Prompt Chaining - AnomalyAnalysis

graph TD
    Start([Workflow Start: SpatialAnomaly]) --> A1[ProcessSensorDataActivity]
    A1 --> Gate1{Data Valid?}
    Gate1 -->|No| Fail([Return Failed Result])
    Gate1 -->|Yes| A2[ClassifyAnomalyActivity]
    A2 --> A3[ScientificAnalysisActivity]
    A3 --> A4[RiskAssessmentActivity]
    A4 --> Gate2{Risk = CRITICAL?}
    Gate2 -->|Yes| Alert[AlertBridgeActivity]
    Gate2 -->|No| A5[GenerateRecommendationActivity]
    Alert --> A5
    A5 --> End([Return AnalysisResult])
    
    style Start fill:#e1f5ff
    style End fill:#d4edda
    style Fail fill:#f8d7da
    style Alert fill:#fff3cd
    style Gate1 fill:#fff3cd
    style Gate2 fill:#fff3cd

Loading

Routing - GalacticAnomalyClassifier

graph TD
    Start([Workflow Start: SpaceAnomaly]) --> A1[ClassifyAnomalyActivity]
    A1 --> Router{Anomaly Type?}
    Router -->|Temporal Rift| A2[AnalyzeTemporalRiftActivity]
    Router -->|Dark Matter| A3[AnalyzeDarkMatterActivity]
    Router -->|Alien Artifact| A4[AnalyzeAlienArtifactActivity]
    Router -->|Stellar Phenomenon| A5[AnalyzeStellarPhenomenonActivity]
    Router -->|Dimensional Tear| A6[AnalyzeDimensionalTearActivity]
    A2 --> End([Return Analysis Result])
    A3 --> End
    A4 --> End
    A5 --> End
    A6 --> End
    
    style Start fill:#e1f5ff
    style End fill:#d4edda
    style Router fill:#fff3cd
    style A2 fill:#cfe2ff
    style A3 fill:#cfe2ff
    style A4 fill:#cfe2ff
    style A5 fill:#cfe2ff
    style A6 fill:#cfe2ff

Loading

Parallelization - StarshipDiagnostics

graph TD
    Start([Workflow Start: Starship]) --> Parallel{Parallel Scans}
    Parallel -->|Parallel| S1[HullIntegrityScanActivity]
    Parallel -->|Parallel| S2[ReactorCoreScanActivity]
    Parallel -->|Parallel| S3[NavigationScanActivity]
    Parallel -->|Parallel| S4[LifeSupportScanActivity]
    Parallel -->|Parallel| S5[WeaponsScanActivity]
    S1 --> Gather[Gather Results]
    S2 --> Gather
    S3 --> Gather
    S4 --> Gather
    S5 --> Gather
    Gather --> Gate{Critical Findings?}
    Gate -->|Yes| Vote{Parallel Voting}
    Gate -->|No| Aggregate[AggregateResultsActivity]
    Vote -->|Parallel| V1[SafetyVoterActivity]
    Vote -->|Parallel| V2[SeverityVoterActivity]
    Vote -->|Parallel| V3[RecommendationVoterActivity]
    V1 --> GatherVotes[Gather Votes]
    V2 --> GatherVotes
    V3 --> GatherVotes
    GatherVotes --> Aggregate
    Aggregate --> End([Return DiagnosticReport])
    
    style Start fill:#e1f5ff
    style End fill:#d4edda
    style Parallel fill:#fff3cd
    style Gate fill:#fff3cd
    style Vote fill:#fff3cd
    style S1 fill:#cfe2ff
    style S2 fill:#cfe2ff
    style S3 fill:#cfe2ff
    style S4 fill:#cfe2ff
    style S5 fill:#cfe2ff
    style V1 fill:#ffd6cc
    style V2 fill:#ffd6cc
    style V3 fill:#ffd6cc

Loading

Orchestrator-Workers - SpaceColonyPlanner

graph TD
    Start([Workflow Start: ColonyRequest]) --> A1[AnalyzePlanetActivity]
    A1 --> A2[DetermineStructuresActivity]
    A2 --> Router{Dynamic Structure Routing}
    Router -.->|HabitatDome| W1[PlanHabitatDomeActivity]
    Router -.->|PowerPlant| W2[PlanPowerPlantActivity]
    Router -.->|Agriculture| W3[PlanAgricultureActivity]
    Router -.->|MiningFacility| W4[PlanMiningFacilityActivity]
    Router -.->|ResearchLab| W5[PlanResearchLabActivity]
    Router -.->|DefenseSystem| W6[PlanDefenseSystemActivity]
    W1 --> Gather[Task.WhenAll - Gather Worker Results]
    W2 --> Gather
    W3 --> Gather
    W4 --> Gather
    W5 --> Gather
    W6 --> Gather
    Gather --> A3[SynthesizePlanActivity]
    A3 --> A4[OptimizeTimelineActivity]
    A4 --> End([Return ColonyMasterPlan])
    
    style Start fill:#e1f5ff
    style End fill:#d4edda
    style Router fill:#fff3cd
    style Gather fill:#fff3cd
    style W1 fill:#cfe2ff
    style W2 fill:#cfe2ff
    style W3 fill:#cfe2ff
    style W4 fill:#cfe2ff
    style W5 fill:#cfe2ff
    style W6 fill:#cfe2ff

Loading

Evaluator-Optimizer - AlienTranslator

graph TD
    Start([Workflow Start: AlienText]) --> Check1{First Iteration?}
    Check1 -->|Yes| A1[TranslateActivity]
    Check1 -->|No| A3[RefineTranslationActivity]
    A1 --> A2[EvaluateTranslationActivity]
    A3 --> A2
    A2 --> Gate1{Quality ≥ 8.0 AND<br/>Meets Standards?}
    Gate1 -->|Yes| Success([Return Success Result])
    Gate1 -->|No| Gate2{Max Iterations<br/>Reached?}
    Gate2 -->|Yes| MaxReached([Return Best Effort Result])
    Gate2 -->|No| Restart[ContinueAsNew<br/>Restart Workflow]
    Restart -.-> Check1
    
    style Start fill:#e1f5ff
    style Success fill:#d4edda
    style MaxReached fill:#fff3cd
    style Gate1 fill:#fff3cd
    style Gate2 fill:#fff3cd
    style Restart fill:#e1f5ff

Loading

Autonomous Agent - SpaceDebrisAgent

graph TD
    Start([Workflow Start: MissionParameters + AgentState]) --> A1[AgentReasoningActivity]
    A1 --> Router{Agent Decision?}
    Router -->|COMPLETE_MISSION| A2[GenerateReportActivity]
    A2 --> End([Return MissionResult])
    Router -->|SCAN_DEBRIS_FIELD| T1[ScanDebrisFieldActivity]
    Router -->|ANALYZE_DEBRIS| T2[AnalyzeDebrisActivity]
    Router -->|MOVE_TO_LOCATION| T3[MoveToLocationActivity]
    Router -->|CHECK_FUEL| T4[CheckFuelActivity]
    Router -->|CAPTURE_DEBRIS| T5[CaptureDebrisActivity]
    Router -->|REQUEST_HUMAN_APPROVAL| T6[RequestHumanApprovalActivity]
    T1 --> Update[UpdateAgentState]
    T2 --> Update
    T3 --> Update
    T4 --> Update
    T5 --> Update
    T6 --> Wait[WaitForExternalEvent: HumanApproval]
    Wait --> Update
    Update --> Check{Failure Conditions?}
    Check -->|Fuel Exhausted| Fail([Mission Aborted])
    Check -->|Max Steps| Fail
    Check -->|Continue| ContinueAsNew[ContinueAsNew: Next Step]
    ContinueAsNew --> A1
    
    style Start fill:#e1f5ff
    style End fill:#d4edda
    style Fail fill:#f8d7da
    style Router fill:#fff3cd
    style Check fill:#fff3cd
    style T1 fill:#cfe2ff
    style T2 fill:#cfe2ff
    style T3 fill:#cfe2ff
    style T4 fill:#cfe2ff
    style T5 fill:#cfe2ff
    style T6 fill:#e7d4ff
    style Wait fill:#e7d4ff
    style ContinueAsNew fill:#fff3cd

Loading

Prerequisites

1. Install a container orchestration tool such as Docker Desktop or Podman.
2. Install .NET 9 SDK
3. Install Dapr CLI
4. Install Ollama
5. Initialize Dapr: dapr init
6. Run Ollama server: ollama serve and run a model: ollama run llama3.2:3b
7. Run Diagrid Dashboard: docker run -p 8080:8080 ghcr.io/diagridio/diagrid-dashboard:latest

See the individual demo README files for specific instructions how to run the applications.

Next Steps

• Join the Dapr Discord to connect with the community.
• Explore the Diagrid Dev Dashboard for workflow observability.
• Run your workflows and AI Agents reliably with Diagrid Catalyst.