BoxLang AI 3.0 introduces a structured multi-agent orchestration system where AiAgent instances form a hierarchy tree. Sub-agents are automatically registered as callable tools via addSubAgent(), enabling LLM-driven delegation without manual wiring. Key architectural changes include fully stateless agents with per-call identity resolution for userId and conversationId, enabling safe concurrent multi-user serving. The post also covers the agent run lifecycle, streaming in multi-agent setups, suspend/resume via checkpointing for human-in-the-loop workflows, and built-in cycle detection to prevent infinite delegation loops.
Table of contents
The Agent TreeSub-Agents as ToolsAiAgent is Now Fully StatelessPer-Call Identity Routing on MemoryThe Agent Run LifecycleStreaming with Multi-Agent TeamsSuspend and ResumeIntrospectionA Complete Multi-Agent ExampleWhat's NextSort: