Structured Visual Agents

Structured Visual Agents are an advanced type of visual agent, composed of a sequence of blocks.

Structured Visual Agents can visually implement complex agentic logic, in a fully transparent and controllable manner, with support for deterministic logic.

Recommendation: Before using Structured Visual Agents, familiarize yourself with Simple Visual Agents.

To inspect a run while it executes, see Agent Run Inspector.

• Overview and example use cases
  • Predefined sequence of steps and branches
  • Gathering “mandatory” information at the beginning of the conversation
• Key concepts
  • Blocks
  • State and Scratchpad
  • Expressions and templating
  • Starting block and subsequent turn behavior
  • Pre-turn and post-turn blocks
• Blocks
  • Agentic Loop
    • Exit Conditions
      • Example use case
    • State and scratchpad virtual tools
    • Forcing tool call arguments
    • Using Agents as tools
    • History passing
  • LLM Request
  • Manual and Mandatory Tool Call
    • Output
  • Delegate to Another Agent
    • Limitation
  • Routing
  • For Each
    • Example use case
    • Output
    • Parallel execution
    • Cautions
  • Parallel
    • Output
    • Parallel execution
    • Cautions
  • Reflection
    • Limitation
  • Set State and Scratchpad entries
  • Context Compression
  • Generate Text Output
  • Generate Artifact
    • Requirement
  • Python Code
    • Special variables
      • State and scratchpad
      • Output of other blocks
      • Conversation history
      • Agent context
  • Long-Term Memory
    • Setup
    • Types of memory items
    • Memory banks
  • Semantic Feedback Management
    • Setup
    • Usage
• Expressions and templates
  • CEL expressions
    • Variables
  • CEL templates
  • Jinja
• How to …
  • Conversational Disambiguation: Force an agent to uniquely identify the subject of discussion before moving on