Nasdanika Templates

Flow

This page explains how agentic flows work in Draw.AI: the simple core model (LLMs + tools + prompts), the swimlane notation used here, the core concepts and contracts, and how diagrams become executable artifacts (Python via CrewAI demos, Java runtimes, JSON/Ecore exports). The examples and patterns below are grounded in the Draw.AI approach and mapping capabilities.

Overview

At its heart, an agentic flow is a sequence of LLM calls and tool calls organized so that agents can collaborate, hand off work, and produce typed inputs and outputs. The flow is a visual, machine‑readable specification that:

  • Defines agents and the tasks they perform.
  • Declares tools (APIs, DBs, microservices, LLMs) and their contracts.
  • Binds inputs, outputs, tool parameters and return values to ontology types.

Key idea: constructs like “agent”, “crew”, or “task” are syntactic sugar — they structure LLM + tool interactions so humans can reason about them. Under the hood, everything is LLMs provided with prompts and tools.

Core Concepts and Definitions

  • LLM - A language model used to interpret prompts, generate text, or orchestrate logic. In flows,
  • Tool - Any callable capability: HTTP API, database query, file store, or a wrapped LLM. Tools have typed parameters and typed return values defined in the ontology.
  • Agent - A logical actor that performs tasks by issuing LLM calls and tool calls. An agent receives typed inputs, uses tools/LLMs, and produces typed outputs. Agents are swimlanes in the notation used on this page.
  • Task - A unit of work inside an agent swimlane. Tasks map to one or more LLM/tool calls and have explicit input/output contracts.
  • Tool Call - A direct invocation of a tool with typed parameters; returns typed results.
  • Agent Call - A logical invocation of another agent. An agent call can be wrapped as a tool call (i.e., the callee agent is exposed as a callable tool), enabling uniform orchestration.
  • Hand‑over - The explicit transfer of an output from one agent/task to another agent/task. Hand‑overs are typed and validated against the ontology.

Swimlane Notation Explained

This page uses a swimlane notation where:

  • Each swimlane represents an Agent (label at top).
  • Inside a swimlane, tasks are boxes connected by arrows.
  • Tool icons sit beside tasks in the same swimlane. Tools can be connected to tasks.
  • Hand‑overs are arrows between tasks with labels referencing ontology types (e.g., CustomerProfile -> RiskAssessment).

Why swimlanes? They align naturally with process thinking and make responsibilities explicit. Other notations are valid — the important part is that the notation maps to the same underlying model: LLM calls + tool calls + typed contracts.

From Diagram to Executable Artifact

Draw.AI flows are designed to be first‑class artifacts that can be exported or executed:

Generation targets

  • CrewAI Python — diagrams can generate executable Python orchestrations (demoed with CrewAI).
  • LangGraph, Semantic Kernel, Java runtimes — flows can be mapped to other runtimes; sometimes you generate code, sometimes you run the flow directly in a runtime that interprets the diagram.

No code required

You do not always need to generate code. Some runtimes can interpret the diagram directly and execute it as an artifact. Other times you generate code for a target platform. Both approaches are supported.

CrewAI Python Demos

There are two functional demos with semantic mapping to the CrewAI model and CrewAI as the generation target:

The demos are dated - they were created as a proof of concept.