//001
Auto-Routing
A fast LLM classifier analyzes each incoming query and automatically routes it to the optimal execution mode: ReAct (Standard) for simple tasks, or DAG (Planning) for complex multi-step workflows. This is the default mode.
Intelligent Task Orchestration
[n.01/03]> Orchestration Engine
Agent doesn't follow fixed scripts. It decomposes tasks on its own, runs multiple steps concurrently, and auto-adjusts plans when problems arise — like a thinking executor, not a button-pressing robot.
[n.02/03]> Capabilities
//002
Standard Mode (ReAct)
Agent thinks and acts step by step. After each step, it decides what to do next based on results. Ideal for tasks requiring judgment and trial-and-error, like alert investigation, document analysis, and information retrieval.
//003
Planning Mode (DAG)
Agent first creates a complete execution plan, decomposing tasks into steps and identifying dependencies: independent steps run concurrently, dependent steps run sequentially. After completion, it checks results and auto-retries with adjusted plans if goals aren't met.
//004
Visual Pipeline (Blueprint Mode)
A full visual workflow editor with 25 node types. Design agent pipelines with drag-and-drop, supporting three progressive levels: Fully Static: Like Dify workflows: each node performs a fixed operation with deterministic I/O connections Semi-Dynamic: Key nodes upgrade to Agent mode, reasoning autonomously within predefined tools and knowledge bases Fully Dynamic: No static pipeline, Agent plans and executes completely autonomously
//005
Extended Thinking
Chain-of-thought reasoning for supported models including OpenAI o-series, Gemini 2.5+, and Claude. The reasoning process is surfaced in the UI, giving users full visibility into how the agent arrives at its conclusions.
//006
Three Model Roles
Assign different models to three distinct roles: General (primary reasoning), Fast (lightweight classification and routing), and Reasoning (deep chain-of-thought). Each role has isolated per-tier environment configuration, allowing you to optimize cost and performance across the pipeline.
//007
Context Management
Long conversations auto-compress earlier content while preserving recent text verbatim. Critical information (task descriptions, user corrections) is protected from compression. A unified token budget manager ensures conversations always stay within the model's context window.
[n.03/03]> Related
/ Related Platform Capabilities /
// [#connect]// [#orchestrate]// [#review]// [#connect]// [#orchestrate]// [#review]// [#connect]// [#orchestrate]// [#review]
// [#connect]// [#orchestrate]// [#review]// [#connect]// [#orchestrate]// [#review]// [#connect]// [#orchestrate]// [#review]
Two ways to run