Chapter 1 Overview

What is Agentish and how does it work?

What is Agentish?

Agentish is a visual workflow editor for building autonomous AI agents. You design your agent workflow by placing nodes on a canvas and connecting them with edges — no code required. When you’re done, Agentish exports your design as a bundle that gets compiled into executable Python code.

Key point: Agentish does not execute agents. It creates workflow definitions. Execution happens in the Agentish sandbox, which compiles your workflow and runs it against challenge environments.

What is a Workflow?

A workflow is a directed graph of AI agents working together. Each agent (node) has a specific role — one might analyze data, another might make routing decisions, another might perform specialized tasks with tools. Edges define the execution order: after one agent finishes, execution flows to the next.

Think of it as an assembly line where each station is an LLM with its own instructions and capabilities, and the product (state) moves from station to station.

Want the theory? For a deep dive into why multi-agent workflows exist and how to reason about their design, see Agentic Workflow — Chapter 1: Foundations.

The Four Node Types

Every Agentish workflow is built from exactly four types of nodes:

Node	Purpose	Quick Rule
Entry Point	Where execution begins. Defines the global state schema.	Exactly one per workflow.
LLM Node	An AI agent that reasons, generates text, and optionally calls tools.	At least one must have no outgoing edge (terminal node).
Router Node	An LLM-powered switch that sends execution down one of N paths.	Must have at least 2 outgoing edges.
Worker Node	A subtask agent that executes tools and returns results to an LLM node.	Must be connected to at least one LLM node.

Each node type is covered in detail in its own chapter (Ch 2–Ch 5).

The Visual Editor

The Agentish editor is a browser-based canvas powered by the LiteGraph library. Here’s how you interact with it:

Action	How
Add a node	Right-click on the canvas → select a node type from the menu.
Connect nodes	Click and drag from an output slot (right side) to an input slot (left side) of another node.
Configure a node	Click on a node to open its property inspector in the right panel.
Delete a node or edge	Select it and press `Delete` or `Backspace`.
Pan the canvas	Click and drag on empty space, or use the scroll wheel.
Zoom	Scroll wheel or pinch gesture.

The Pipeline: Design → Export → Compile → Run

The Pipeline: Visual Editor → Bundle (ZIP) → Compiler → Runner

Design — Build your workflow in the visual editor.
Export — Click “Download Bundle” to get a ZIP containing asl.json (your workflow) and layout.json (visual positions for re-importing).
Compile — Upload the bundle to the sandbox. It validates your ASL, checks for topology errors, and generates executable LangGraph Python code.
Run — The sandbox spins up Docker containers for your agent and the challenge’s MCP tool servers, executes your agent, and streams back logs and results.

Validation happens at export time too. When you click “Download Bundle,” Agentish runs topology checks on your graph before creating the ZIP. If there are errors (e.g., no terminal LLM node, orphan nodes, invalid loops), the download is blocked and you’ll see error toasts explaining what to fix.

Two Edge Types

Edge Type	Created By	Meaning
NormalEdge	Connecting any non-Router node to another node.	“After A finishes, always go to B.”
ConditionalEdge	Connecting a Router node to its targets.	“After Router decides, go to whichever target matches the decision.”

Chapter Summary

Key Takeaways:

Agentish is a visual editor that produces workflow definitions — it does not execute agents.
Workflows are directed graphs built from four node types: Entry Point, LLM Node, Router Node, Worker Node.
The pipeline is: design in editor → export bundle → upload to sandbox → sandbox compiles and runs.
Topology validation catches structural errors before you ever run your agent.

Chapter 2: The Entry Point →