AI Agents Fundamentals in 21 Minutes
AI agents are iterative workflows where AI breaks tasks into steps, uses tools, and refines outputs—unlike simple one-shot prompting. Four core patterns exist: reflection, tool use, planning, and multi-agent systems. Multi-agent architectures (sequential, hierarchical, hybrid, parallel, asynchronous) enable specialized agents to collaborate. No-code tools like n8n let you build functional agents without coding. The biggest opportunity: every SaaS company will have an AI agent equivalent.
What Is an AI Agent
Non-Agentic vs. Agentic Workflows
A non-agentic workflow is one-shot prompting: ask AI to complete a task from start to finish in one go. An agentic workflow breaks the task into iterative steps—outline, research, draft, revise, repeat—creating a circular process that continuously improves output until reaching the final result.
Three Levels of AI Capability
Level 1 is non-agentic (direct request). Level 2 is agentic workflows with iterative steps but human or system guidance. Level 3 is fully autonomous AI agents that independently determine steps, select tools, and refine outputs without external direction—currently not yet achieved at scale.
Four Agentic Design Patterns
Reflection
Ask an AI to review its own output for correctness, style, and efficiency, then provide constructive criticism to improve it. This can be extended by having a second AI prompt the first to perform reflection, creating a multi-agent framework.
Tool Use
Equip AI with tools like web search, code execution, object detection, email access, or calendar scheduling. This allows AI to break down tasks and execute specific parts more effectively—e.g., searching reviews before summarizing, or running code to calculate compound interest.
Planning and Reasoning
AI receives a task and independently determines the exact steps and tools needed to accomplish it. For example, generating an image matching a pose, then describing it via text-to-speech requires sequencing multiple specialized models (pose detection, image generation, transcription).
Multi-Agent Systems
Instead of one AI handling everything, deploy multiple large language models with different specialized roles. Research shows multi-agent results are generally better than single-AI approaches, similar to how specialized teams outperform generalists on complex projects.
Memory Aid: Red Turtles Paint Murals
Mnemonic for the four patterns: Reflection, Tool use, Planning, Multi-agents (RTPM).
Multi-Agent Design Patterns
Single Agent Building Block
Every agent has four components: Task (what it should do), Answer (desired output), Model (LLM choice), Tools (resources available). Mnemonic: TAMT (Tired Alpaca's Mix Te). Example: travel planner agent with task 'plan 3-day Tokyo trip on budget', answer 'detailed itinerary with costs', model Claude, tools Google Maps/Skyscanner/Booking.com.
Sequential Pattern
Simplest multi-agent design: one agent completes a task, passes output to the next agent, and so on—like an assembly line. Example: document processing where agent 1 extracts text, agent 2 summarizes, agent 3 extracts action items, agent 4 saves to database.
Hierarchical Pattern
A manager or leader agent supervises multiple specialized sub-agents, each with their own task. Sub-agents report results back to the manager, who compiles and synthesizes findings. Example: business report where manager delegates to market-trend agent, customer-sentiment agent, and internal-metrics agent.
Hybrid Pattern
Combines sequential and hierarchical structures with continuous feedback loops. Common in robotics and adaptive systems where top-level planning must continuously adjust based on real-time sub-agent feedback. Example: autonomous vehicle with route-planning agent receiving continuous input from sensor-fusion, collision-avoidance, and road-condition agents.
Parallel Pattern
Multiple agents work on different chunks of a task simultaneously to speed up processing, then merge results. Common in large-scale data analysis where different agents process separate data segments independently.
Asynchronous Pattern
Agents execute tasks independently at different times, proving more resilient to uncertain conditions than sequential or parallel approaches. Example: cybersecurity threat detection with agents monitoring network traffic, suspicious patterns, and random usage tests independently, flagging anomalies as they occur.
Complexity and Chaos Trade-off
As multi-agent systems grow more complex with more agents and interconnections, internal chaos increases because there is no direct control over agent interactions. This mirrors how larger human companies become more chaotic and require stronger hierarchies and organizational structures.
Building AI Agents: No-Code Example
n8n Telegram AI Assistant Workflow
A no-code tool (n8n) can build functional multi-agent systems. Example: Inky Bot is a Telegram-based AI assistant that accesses Google Calendar to show daily tasks and can create calendar events. It accepts both text and voice input (voice is transcribed via OpenAI first), then routes to an AI agent with GPT-4o mini, task definition, and calendar tools.
Agent Configuration in n8n
The Inky Bot agent uses TAMT: Task is user query about daily priorities; Answer is prioritized to-do list plus scheduled calendar events; Model is OpenAI GPT-4o mini (swappable); Tools are get-calendar-events and create-calendar-events. No coding required; all configured visually.
Opportunities and Guidance
The SaaS-to-AI-Agent Thesis
Y Combinator guidance: for every SaaS (software-as-a-service) company that exists today, there will be a corresponding AI agent company. This is the single most actionable piece of advice for deciding what to build. Examples of SaaS companies with potential AI agent equivalents: Adobe, Microsoft, Salesforce, Shopify, Linktree, Canva, Squarespace.
Prompt Engineering as High-ROI Skill
AI agents and agentic workflows rely heavily on prompt engineering. Learning to craft effective prompts is one of the highest-ROI skills today because it directly multiplies productivity and output quality across all AI tool usage.
Notable quotes
For every software as a service company there will be a corresponding AI agent company. — Y Combinator (cited by Tina Huang)
An agentic workflow is more a circular iterative process: think, do research, come up with output, revise, repeat. — Tina Huang
As you increase the complexity of these systems you're also increasing the amount of chaos within it. — Tina Huang
Action items
- Identify a SaaS company you know well and sketch out what its AI agent equivalent would do and how it would differ.
- Build a simple no-code AI agent workflow using n8n or Make.com to automate a personal task (e.g., calendar management, email triage, research compilation).
- Practice prompt engineering using the TAMT framework (Task, Answer, Model, Tools) to define clear agent instructions.
- Experiment with multi-agent patterns by creating two or more agents with different roles and testing how they collaborate on a shared task.