Workflows in AutoGen Studio (Experimental)

How it Works: Example Story

flowchart TD
    subgraph Backend
        subgraph Workflow
            Step1["Step 1: Fetch Web Page"]
            Step2["Step 2: Summarize Content"]
            Step3["Step 3: Format Output"]
            Step1 --> Step2 --> Step3
        end
        Runner["Workflow Runner"]
        Runner --> Workflow
    end
    UI["UI (Frontend)"]
    User["User"]
    User --> UI
    UI -- "Builds & Starts Workflow" --> Runner
    Runner -- "Streams Progress & Results" --> UI
    UI -- "Shows Results" --> User

A user builds a workflow in the UI (e.g., "Summarize a web page"). The UI sends the workflow to the backend runner, which runs each step (fetch, summarize, format). Progress and results stream back to the UI for live visualization.

Architecture Overview

flowchart TD
    UI["UI (Workflow Builder & Visualizer)"]
    API["API (REST & WebSocket)"]
    Workflow["Workflow (Container)"]
    Step["Step (Unit of Computation)"]
    Edge["Edge (Transition/Condition)"]
    Runner["Runner (Executes Workflow)"]
    State["State (Shared/Step State)"]
    Serialization["Serialization (JSON Schema)"]
    Streaming["Streaming Events (WebSocket)"]

    UI -- "Author/Modify Workflows" --> API
    API -- "Serialize/Deserialize Workflows & Steps" --> Serialization
    API -- "Start/Stop/Update Workflows" --> Workflow
    Workflow -- "Contains" --> Step
    Workflow -- "Contains" --> Edge
    Workflow -- "Holds" --> State
    Workflow -- "Is Executed By" --> Runner
    Runner -- "Runs Steps (Parallel/Sequential/Conditional)" --> Step
    Edge -- "Defines Transition/Condition" --> Step
    Runner -- "Streams Events" --> Streaming
    Streaming -- "Updates" --> UI
    Step -- "Defines Input/Output Types (Schema)" --> Serialization
    Serialization -- "Enables UI Authoring & Validation" --> UI
    State -- "Shared/Forwarded Between Steps" --> Step
    Runner -- "Updates State" --> State
    API -- "WebSocket for Real-Time Updates" --> Streaming
    UI -- "Visualizes Status/Progress" --> Streaming

What is a Workflow?

A Workflow is a container for a set of Steps (units of computation) and Edges (transitions/conditions between steps). Workflows enable deterministic, composable chains of logic, supporting parallel, sequential, and conditional execution.

Steps: Each step defines input/output types (with full schema), is fully serializable, and can be validated. Steps can transform data, call APIs, or run agents. (code)
Edges: Define transitions between steps, with optional conditions (output, state, or expression-based). (code)
Workflow Class: Holds steps, edges, state, and metadata. Supports serialization for UI authoring. (code)
Runner: Executes workflows, supporting parallel, sequential, and conditional logic. Streams real-time events for UI updates. (code)

UI & Serialization

Serialization: All steps and workflows are defined with JSON schemas, enabling UI-based authoring, validation, and modification.
UI Integration: The UI allows users to build, edit, and visualize workflows. Real-time status and progress are streamed via WebSocket events. (API code)

Agents as Steps (Anything as Steps)

Custom Steps: Any agent or logic can be added as a step by implementing the BaseStep interface. Steps must define input/output types as Pydantic models for validation and UI compatibility.
Examples: See AgentStep for LLM agents, FunctionStep for arbitrary functions, and HttpStep for HTTP calls.
State Access: Steps read/update workflow state via the provided Context object (context.get() / context.set()).
Requirement: All steps must specify input/output schemas and implement the execute(input_data, context) method.

Example Workflows

For more details, see the core implementation, step definitions, and UI API integration.

4.4 KiB Raw Blame History