agent-orchestration

A coordinator pattern dispatches tasks to specialist agents based on request type, improving clarity and debuggability. It applies to complex workflows like design coordination and code generation where dynamic tool selection and session state are important.

Pass context between multiple LLM agents using lightweight session state for inter-agent communication. This enables coordinated workflows by maintaining shared context across dynamic multi-stage pipelines and specialist agent dispatches.

Wrap specialist agents as callable tools to compose hierarchical, reusable capabilities. This agent-as-tool pattern allows a root coordinator to delegate complex subtasks dynamically while maintaining structured routing and stateful collaboration.

LlmAgent provides dynamic reasoning for adaptable tasks, whereas WorkflowAgent executes deterministic flows for predictable processes. Selecting between them depends on whether your multi-agent orchestration requires flexible reasoning or strict process execution.

Avoid monolithic agents and circular dependencies when building multi-agent systems. Use structured coordinator-based routing and anti-pattern guidance to ensure modular, maintainable, and scalable agentic systems with robust error handling.

Implement advanced orchestration patterns with robust error handling to manage failures in multi-stage agent pipelines. This ensures production workflows remain stable when dynamic selectors or specialist agents encounter execution errors.