Mastering
AI Agents
in 60 min
From Zorro Cheng · 2026
Introduction
Context, velocity, and why 2026 is different
What will happen to AI Agents in 2026?
Time it took for platforms to reach 1 million users
AI Agent is here now!
No/Low code programming is fading
Traditional drag-and-drop coding approaches are becoming obsolete. Agents write and deploy code on your behalf — no GUI builder required.
RPA is dying
Coding Agents and Auto Browsers are transforming automation. They reason, adapt, and handle exceptions — not just replay fixed scripts.
Diffusion of Innovation
AI Ecosystem
Picking the right AI for the job
Sport Shoes Shopping
Brands → Products → Stores (many-to-many connections)
AI Shopping 2026
Same structure — Vendors → Models → Surfaces
Sport Shoes 2026 — 4 Tiers
Adding the fourth tier: in-store experience zones
AI Shopping 2026 — 4 Tiers
The orange row = products/agents we'll come back to
Notice the orange row — these are where agents live. We'll come back to them in Half 2.
One Model CAN'T Fit All
Build your AI toolbox — each model excels in unique ways
Gemini Pro
Multimodal champion with exceptional visual understanding and native Google Workspace integration
OpenAI GPT-5
Thoughtful problem-solver with strong analytical and reasoning capabilities across domains
Claude Opus
Go-to for complex coding, long documents, and nuanced instruction-following
DeepSeek
Specialized in Chinese-language tasks and cultural contexts; strong coding and reasoning
Understanding AI Model Fundamentals
When evaluating models, check these 2 critical attributes
Smartness
Accuracy without hallucination. How often does the model produce correct, reliable answers? Does it make up citations, invent facts, or confidently say wrong things?
Memory Capacity
Context Window — how much text the model can process at once. Bigger = can handle longer documents, longer conversations, more context.
Understanding Tokens
Example
1 token ≈ 4 characters · ¾ of an English word
Cost
APIs charge per token consumed — both input (prompt) and output (response)
Includes
Everything in your conversation: your messages, AI replies, documents pasted in
Context Window
The size of the AI model's working memory — all it can "see" right now
Subscription vs API
🍱 Subscription — "All You Can Eat"
- Flat ~US$20/month (ChatGPT Plus, Claude Pro, Gemini Advanced)
- Usage quota — resets on a rolling window
- Predictable cost, hits a wall when quota exceeded
- Best for: individuals, daily chat, exploration
🔌 API — Pay As You Go (per token)
- Claude Sonnet: ~$3/1M input tokens, ~$15/1M output
- Sample Q&A: 10K input = $0.03 + 2K output = $0.03 → $0.06/req
- No cap — cost scales with usage linearly
- Best for: apps, automation, agents, variable workloads
When you move from chat to agents, you almost always move to API — agents burn tokens fast. This is the cost shift to expect in Half 2.
AI Memory = Context Window
- No long-term memory between sessions — each new chat starts fresh
- Within a session, the context window IS its working memory
- When the window fills up, the AI starts forgetting the earliest parts
- Bigger model = bigger memory (Gemini Pro 1M, Claude 200K)
- Long chats need handover summaries — same as a temp worker
AI Model Memory Comparison
| Model | Context Window | Best Use Case |
|---|---|---|
| Gemini Pro | 1,000,000 tokens | Multimodal tasks, long-context coding |
| LLaMA | 10,000,000 tokens | Open-source applications, efficient processing |
| GPT-5 | 400,000 tokens | General purpose, multimodal processing |
| Claude Sonnet/Opus | 200,000 tokens | Document analysis, research, complex reasoning |
| DeepSeek | 128,000 tokens | Code generation, reasoning, large-context understanding |
Larger context = more working memory per session, but usually higher cost per API call.
How to Collaborate
Talking to AI effectively — the mindset shift
Think of AI as your
MIT Graduate Intern
Brilliant — but new on the job.
Working with Your AI Intern
Smarter model, more capable intern
Advanced models handle complex tasks with less supervision. Upgrade your model when you need more reliability.
Set appropriate expectations
Understand strengths and limitations. AI excels at drafting, analysis, and synthesis — but verify anything critical.
Clear, specific instructions
Examples yield far better results than vague descriptions. The more specific you are, the less revision you'll need.
Build productive workflows
Systematic processes that leverage AI strengths while keeping humans in the review and decision loop.
The Temporary Worker Model
AI chats are like a temp worker
Limited memory, starts fresh each session, can only see the current conversation
Provide clear goals upfront
Well-defined objectives help AI understand purpose from the very start
Define a clear job scope
Setting boundaries helps AI stay relevant and avoid scope creep
Ask for a handover document
Request summaries to preserve important information before starting a new session
The context window determines memory
Everything the AI can reference is what you've put in this conversation window
CAST: The Prompt Framework
C — Character or Target Audience
Define who the AI should be, or who the output is for.
"Act as an experienced data scientist..."
A — Aim or Goal
Clearly state what you want to accomplish.
"Create a marketing plan for..."
S — Specific Detail or Context
Provide industry context, constraints, and background — the more the better.
T — Template or Format
Specify how you want the answer presented.
"As a bulleted list," "In a table," "Under 200 words"
Using this structured approach helps your AI "intern" deliver exactly what you need — first time.
Learning Prompting from AI
Let AI teach you
Ask AI to analyze your prompts and suggest specific improvements to structure and clarity
Ask AI to write a prompt for you
For complex requests, ask AI to formulate the optimal prompt structure before you run it
Especially useful for image AI
Image models (Midjourney, DALL·E, Imagen) have very specific prompt templates — ask them first
Build a prompt library
Identify patterns in successful prompts and save them as reusable templates in your workflow
From LLM Chat
to Agent
Everything before: AI you talk to.
Everything after: AI that does.
Chat → Agent
💬 Chat
- Single turn: one question, one answer
- You copy-paste the result manually
- You close the loop yourself
🤖 Agent
- Multi-step: AI loops until done
- AI takes real actions (web, files, email)
- You review outputs, not every step
An agent = LLM + tools + a loop, working toward a goal.
The Agent Recipe
LLM
The brain — decides what to do next at each step
Tools
The hands — web, files, code, APIs, databases
Loop
The persistence — plan → act → observe → repeat
Agent Tools — Connectors
Give AI the tools they need to do your job — Gemini Enterprise (April 2026)
💬 Microsoft Teams
Chat and notify across enterprise channels
📁 Google Drive
Read, create, and organize documents in the cloud
🎫 Jira Cloud
File, update, and comment on tickets
📄 Confluence Cloud
Create and edit wiki pages and documentation
🧠 Notion
Manage databases, pages, and team knowledge bases
📧 Microsoft Outlook
Send email and schedule meetings on your behalf
These 6 are among the most-used Gemini Enterprise connectors. They connect agents to where your work actually lives.
Consumer Agents in 2026
Cursor
Coding agent built into your editor. Writes, reviews, and refactors code with full codebase context.
Claude Desktop
Anthropic's desktop agent for chat + tools. Can access files, run code, use MCP connectors.
Manus Desktop
Research and multi-step agent. Browses the web, synthesizes findings, produces documents.
Perplexity Computer
Browser-based agent that surfs and acts. Fills forms, extracts data, completes web-based tasks.
Enterprise Agents — Governance, Compliance, Scale
NotebookLM Enterprise
Secure document agents over your company's knowledge corpus
Gemini Enterprise
Google's agent platform with admin controls and full audit logging
Claude for Enterprise
Anthropic's enterprise tier with SSO, data isolation, and usage controls
Microsoft Copilot Studio
Build, govern, and deploy custom agents across the Microsoft 365 stack
Same agent capabilities — plus governance, compliance, identity management, and audit trails.
Your AI Mastery Journey — 5 Levels
Now that you've met agents, here's how far this goes
I held this map back until now — until you'd met an agent, the higher levels would have been abstract.
How We Engineer with AI — 4 Years, 4 Eras
Engineering
Coding
Engineering
Engineering
Each year, the lever moves higher up the stack. The frontier today is Harness — building agent infrastructure.
Agent Team
Orchestrated AI workforce — specialization + parallel execution
Solo Agent → Agent Team
Solo Agent 😓
One brain doing everything — slow, context-overloaded, generalist output
research + write + review + publish
Agent Team 🚀
One orchestrator + many specialists — parallel, focused, production quality
Same job, divided across roles — like turning a freelancer into a startup.
Skills — SOPs for Your AI
Capability Uplift
Give the agent a skill it's currently weak at. Example: a Skill that makes any agent follow your company's frontend design system precisely.
Encoded Preference
Lock in your workflow your way. Example: a Skill that ensures every report follows your exact structure, tone, and approval chain.
Claude Agent-Team — Orchestration Built In
🎯 Orchestrator + Sub-agents
Main agent delegates tasks to specialist sub-agents, then synthesizes their outputs
🔧 Skills attached per role
Each sub-agent gets the right SOPs — the researcher gets a web-search skill, the writer gets your brand voice
✨ Adhoc agent creation
Spawn new specialists on the fly when the task requires an unexpected capability
Don't define teams up front — let them form for the task at hand.
Metaskills — Skills That Build Skills
Skill Builder
A metaskill that generates a new Skill from your plain-language description. You describe the SOP; the agent writes and tests it.
Team Composer
Analyzes a job you describe and spawns the optimal agent team with the right Skills already attached to each role.
Paperclip AI — Level 4 in Production
- One product, multiple specialist agents (research / draft / publish)
- Orchestrator routes each user request to the right sub-agent team
- Adhoc agents spawned for unfamiliar or complex tasks on the fly
Agent Corp
AI as an organization — multiple teams, multiple orchestrators, one mission
From Team to Organization
Agent Team (Level 4)
One orchestrator, many sub-agents, one mission. Like a startup with a CEO and functional staff.
Agent Corp (Level 5)
Many orchestrators, many teams, running an entire organization. Each team is a department.
Each "team" is a department. Together they form a company.
ZorCorp — AI-First Operating Model
The "company" runs on agents. Humans steer.
You Just Climbed 5 Levels in 60 Minutes
- Chat — you talked to AI
- Tools — you saw the ecosystem & how to pay
- Agent — you understand autonomy
- Agent Team — you saw orchestration & Skills
- Agent Corp — you see the destination
It's "Which level am I building toward?"