Everyone is solving how agents run.
Nobody is solving how agents think.
A mental model for AI Agents that anyone can understand — no code required.
You might be thinking about Agents wrong
In the past year, everyone has been talking about AI Agents. LangGraph, AutoGen, CrewAI — framework after framework, tutorial after tutorial, all telling you how to “run” an agent.
But before you start building, there's a more fundamental question nobody helps you answer:
What is an Agent, really?
Not “an AI system that can autonomously complete tasks.” But: what is its internal structure? How do you design it? How do you communicate what's in your head to an engineer?
Starting from Graph Theory
All complex things are graphs underneath. Graph theory is one of the simplest concepts in mathematics:
City maps are graphs. Social networks are graphs. Your org chart is a graph. Graph theory doesn't care what's inside the nodes — only how they connect.
Graphs + Rules = FSM
A Finite State Machine (FSM) is a graph with two constraints:
Nodes become States
Finite in number — you define all possible states upfront. No surprises, no emergent states.
Edges become Transitions
Conditional — you only move from one state to another when a specific condition is met.
Example: Traffic Light FSM
Three states. Three transitions. Looping forever. The core value of FSM is predictability.
An Agent is an FSM, but the LLM decides the transitions
An AI Agent is essentially an FSM with one difference: transition conditions are decided by the LLM's output, not hardcoded.
Traditional FSM
if user clicks “confirm”
Agent
if LLM judges “task complete”
if LLM judges “need more info”
Graph Theory — structure
→ FSM — graph + deterministic rules
→ Agent — FSM + LLM decides transitions
Prompt is the First-Class Citizen
If an agent is an FSM, what goes inside each state? A Prompt. Each state corresponds to a prompt that defines what to do, how to think, what format to output.
Designing an agent is designing a state graph of prompts.
Two Fundamental Patterns
All agents reduce to two basic structures, or a combination:
Pipeline
Fixed sequence. A's output feeds B, B feeds C. Good for deterministic flows.
Orchestrator + Workers
↑ results flow back
Main prompt has global vision, delegates to specialized workers. Good for dynamic flows.
The fundamental difference: who makes decisions. In a pipeline, each prompt only knows its next step. An orchestrator has global vision and coordinates everything.
The Design Layer is Empty
Every agent tool on the market solves the same problem: how to run an agent.The design layer — the questions you need before writing code — is blank:
How many prompt states does this agent have?
How does information flow between prompts?
What conditions trigger a jump forward, a loop back?
Where are the loops, and when do they terminate?
What tools does each prompt node have access to?
Agent Architect fills this gap.
A canvas where you drag prompt nodes, define transition conditions, annotate tool mount points — and export your design as YAML or JSON.