How to Build AI Agents: 3 Routes, from Zero Code to Full Stack

“I use ChatGPT every day. I’ve tried Claude Code too. What should I do next?”

I’ve been getting this question more and more. The answer is straightforward: step up from “using” to “building.” That’s all it takes.

But the moment you search “how to build an AI agent,” you’re buried in technical articles dense with Python code. In the past two months, I’ve watched more than twenty people quit reading within three minutes.

The reality: building AI agents breaks into three routes. No-code, low-code, full code. Which one you pick isn’t decided by “can you write code?” It’s decided by “what needs to be running in one week for this to count as a win?” If you don’t know this, you’ll take the long way around. Every time.

Why Learn to “Build” in May 2026?

The air around AI agents shifted noticeably when 2026 began.

The signal: Uravation CEO Satoru Sato’s book Claude Jitsumujutsu (Claude Work Skills) went up for preorder in May, with a July 16 release date (hanmoto.com ISBN 9784815645304). A Claude Code business book entering major distribution. This means the “Claude Code = developer tool” label has fully peeled off. It’s now mainstream business infrastructure.

The same Uravation released a May 2026 survey of 25 executives, where the top use case — at roughly 40% — was automated generation of board meeting materials. Not writing code. Materials that everyone prepares every week.

Microsoft moved in May too, officially publishing guidance on “agent governance, intelligent workflows, and connected app experiences” (Microsoft Copilot Studio Blog, https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/new-and-improved-agent-governance-intelligent-workflows-and-connected-app-experiences/). A declaration that AI agent operations belong in enterprise infrastructure, not in individual experimentation. Anthropic and AWS are moving in the same direction.

What happens next: the gap between “users” and “builders” widens at three times the 2025 pace.

Someone who just uses ChatGPT every day is in the same position they were six months ago. Someone who spent those six months running three AI agents now processes three times the workload. That’s the reality of May 2026.

I don’t write scare pieces. But people who stay at the “user” level are being left behind — this is just a fact. Check your current position against five questions.

In the past 30 days, have you sent the same instruction to ChatGPT or Claude five or more times? Is there a recurring weekly task you still process by hand? Can you name three things you’d like to hand to an AI — right now? Have you ever paid more than $10/month for a no-code tool? Have you ever built something that runs on its own — even once?

If three or more answers are “no,” you’re still on the user side. The next hour is how you change that.

[[IMG:3]]

Route 1: No Code — The Fastest Path to a Running Agent

The first route requires zero lines of code.

Target: anyone with no programming background, or anyone whose most recent coding memory is quitting a tutorial site. Doesn’t matter. The no-code AI agent infrastructure as of May 2026 is remarkably mature.

Five representative tools: Claude Projects, ChatGPT GPTs, Notion AI, Zapier (AI Actions), Dify. All available in free tiers or under $20/month. The shared mechanism: define behavior in natural language, templatize it, call it when needed.

Building comes down to four steps.

Step 1: Write the goal in one sentence. “Take a meeting notes text and produce a summary, then post it to Slack.” That level of specificity. “Streamline operations” doesn’t count. You need a clear verb and a clear object. Without these, everything downstream drifts.

Step 2: Define input and output format. Input: meeting notes text. Output: 400-character summary plus attendee list. When the format is fixed, the AI’s variance drops dramatically.

Step 3: Write the instruction (the prompt). In Claude Projects, use “Custom Instructions.” In GPTs, use “Instructions.” Write something like: “You are a meeting notes summarization expert. Follow these rules when summarizing: …”

Step 4: Test and refine. Run three real meeting notes through it. Review outputs. Tighten the instructions. Two rounds of this and you’re at production-ready quality.

I built a meeting-notes-to-Slack notification agent in 30 minutes. Set up a template in Notion AI, wired a trigger in Zapier. Cut two hours of monthly work. That’s 24 hours back per year.

The limits are also clear. Route 1 struggles with complex conditional branching and fine-grained external API integration. A workflow like “read an invoice as PDF, classify accounts, auto-register in accounting software” doesn’t complete in Route 1.

Another case it doesn’t fit: when processing speed matters. Route 1 relies on manual prompt submission or Zapier/Notion workflow triggers — not built for high-volume batch processing. “Triage 100 support tickets in under a second” is Route 2+ territory.

Your action today: write down one task you want to automate, then decide before end of day whether to build it in Claude Projects, GPTs, or Dify. Once decided, it’ll be running tomorrow.

Route 2: Low Code — Hybrid with Claude Code and Cursor

When Route 1 isn’t enough, Route 2 is next.

Target: people who can read code but are afraid to write it. Or people who need capabilities Route 1 can’t handle — connecting multiple APIs, processing internal files, scheduling automation.

The anchor tool here is Claude Code. Claude Code is an AI agent that runs in your terminal. Tell it in natural language: “write a script that sends a Slack notification” — it generates the Python, runs it, and tests it. Cursor is the same family: AI embedded in a VS Code-style editor.

“Instruct in natural language → AI generates code → human reviews.” Repeat that loop. You’re not writing code; you’re reviewing code. That’s the essence of Route 2.

Especially important here: MCP (Model Context Protocol). MCP is an open standard from Anthropic for AI agents to communicate with external tools (Anthropic docs: https://docs.anthropic.com/en/docs/mcp). As of May 2026, major services including Notion, Slack, and GitHub have published MCP implementations. Stand up one MCP server and Claude Code can operate external services directly. What required Zapier as middleware in Route 1 becomes a direct connection in Route 2.

A personal example. Last month I told Claude Code: “Read my Notion meeting notes page, extract the attendee list, and write a script that posts ‘Here are today’s notes’ to the relevant Slack channel.” A working script was complete in 10 minutes. Cron scheduled for every Friday at 6pm: done 30 minutes after I started.

[[IMG:2]]

Anthropic itself demonstrated a related capability in May 2026, publishing a case study deploying AI agent packages across three industries in eight days. That story is covered in The structure behind Anthropic’s 8-day three-industry rollout. That kind of speed doesn’t come from writing everything from scratch. It’s the output of tools like Claude Code used at full throttle.

Route 2’s limits: some code review ability is required. Pushing generated code to main without understanding it causes production incidents. At minimum, you need enough of a trained eye to distinguish “this rewrites a file” from “this calls an external API.”

Another case it doesn’t fit: workflows that need multiple people. Route 2 is primarily local-environment development. Building something “the whole team uses daily” requires additional design work. Route 2 excels for one-person automation. For team-shared infrastructure, Route 3’s architecture thinking becomes necessary.

Your action today: install Claude Code and say “Hello, agent.” The prerequisites on pricing and usage are covered in Claude Code essentials. Read Claude Code pricing: 3 scenarios first if you want to approach it with no surprises.

Route 3: Full Code — Framework Adoption for Production

The third route is full-code, full commitment.

Target: working engineers, or people who have shipped production code before. Or anyone thinking “I want to sell AI agents as a product” or “I want to embed this in company infrastructure.”

Major frameworks: Claude Agent SDK (Anthropic official), LangChain, AutoGen (Microsoft), CrewAI. Each has a different strength. Claude Agent SDK is cleanest for single-agent implementations. LangChain has the broadest external tool integration range. AutoGen excels at multi-agent coordination — multiple AIs conversing while progressing through tasks. CrewAI leans into a role-definition organizational model.

The most important decision in Route 3: architecture selection. Single-agent (one agent handles everything), multi-agent (multiple AIs divide the work), Plan & Execute (a planner and an executor as separate roles). The choice changes development time and running costs by a factor of 2–5x.

Route 3’s evaluation criteria aren’t “does it work” — they’re “does it stay stable,” “can I observe it,” and “can I forecast costs.” Route 2 can also build things that work. What justifies Route 3 investment is: when it breaks, you can find the cause; monthly API costs are predictable; errors surface as alerts. That’s what makes something embeddable in business infrastructure.

Key development here: AWS Kiro, announced May 22, 2026. Kiro is a “spec-driven AI development tool” — marking a shift from momentum-driven “vibe coding” to specification and discipline (Source: XenoSpectrum, May 22, 2026). If you’re heading to Route 3, building in Kiro’s thinking from the start will save time in the long run.

My current system — the Izumo system — runs five content-production AI agents (the Hashira) alongside review, research, and marketing analysis agents in a multi-agent architecture. Built on Claude Agent SDK, with distinct roles and permissions per agent. Codebase: approximately 8,000 lines. Monthly API cost estimate: around ¥60,000. That’s what Route 3 looks like at realistic operating scale.

Route 3’s realistic note: the moment you move to full code, operational responsibility kicks in. Error monitoring. Cost management. Security reviews. You will spend more time on operational design than on writing code. If you can’t commit to that, Route 3 stalls.

Cases it doesn’t fit: “I just want to try something and see” or “I only need one process automated.” Route 3 is complete overkill for that stage. Without team development experience, maintenance costs will exceed expectations. The practical path: get one agent to production in Route 2, then migrate to Route 3 once you’ve identified the gaps.

Your action today: pick one of the four frameworks and run the official tutorial’s sample code. Anthropic’s “Claude Agent SDK Quickstart” is the least friction-prone entry point available right now.

The 4-Point Selection Table and Three Classic Failure Modes

Looking at the three routes lined up, it’s tempting to read them as “Route 1 < Route 2 < Route 3,” where higher is always better. That’s wrong. These are purpose-specific options.

[[IMG:1]]

Four checks to avoid picking wrong.

Check 1: Goal clarity. Can you write in one sentence what you’re automating? If no — start with Route 1. If yes — consider Route 2+.

Check 2: Code experience. Never written a line? Route 1. Can read but afraid to write? Route 2. Written and shipped? Route 3.

Check 3: Budget. Under $10/month? Route 1. $30–$100/month? Route 2. $300+/month acceptable? Route 3 becomes viable.

Check 4 — most important: What needs to be running in one week? One process working with a button press? Route 1. Multiple processes connected and running? Route 2. Company-wide infrastructure other team members actively use? Route 3.

Three failure patterns I see frequently:

Failure 1: Aiming for Route 2, ending up stopped at Route 1. Opening Claude Code but spending the whole day optimizing Notion AI prompts. Tool became the goal. Fix: decide which specific task should be running by end of week, and come back to that target every time you drift.

Failure 2: Jumping to Route 3 immediately. Common trap for engineers. Reading Claude Agent SDK tutorials, touching LangChain samples, tracking AutoGen papers — three weeks pass and nothing is running in production. Crushed by complexity. Fix: honor the sequence: build something in Route 2, then migrate up.

Failure 3: Starting without a clear goal. “I want to make things more efficient” or “I want to do something with AI” — at this level of vagueness, no route leads to completion. Fix: write the one-week finish line in words and set it as your laptop wallpaper. If you’re serious, actually do this.

I’ve personally experienced all three failures. I spent a month tinkering with Claude Code thinking I was in Route 2, and nothing ran. I spent three weeks deep in LangChain documentation thinking it was Route 3, only to realize it wasn’t necessary for my use case. That’s exactly why I can say with confidence: four minutes on the checklist is not a cost to skip.

The 1-Hour Roadmap for Today

For anyone reading this who’s about to say “I’ll start tomorrow” — tomorrow is too late. Do this in the next hour.

[[IMG:4]]

First 15 minutes: Goal setting. Phone notes app or paper — write one line: “When I give it X, it produces Y as output.” Example: “When I paste meeting notes, it outputs a 400-character summary and attendee list.” That’s the complete target. Can’t write it? Ask ChatGPT: “List 3 tasks I might want to automate, and rewrite each as a one-sentence AI agent goal.” You’ll have options in five minutes.

Next 15 minutes: Route selection. Run the four checks. Zero code experience → Route 1 confirmed. Need multiple processes connected in one week → Route 2. If you’re stuck, always default to Route 1. Building up from a working system beats starting from too high.

Next 15 minutes: Tool setup. Route 1: create a Claude Projects account (free). Route 2: install Claude Code (official docs, about 10 minutes). Route 3: clone the chosen framework’s official repo. If installation stalls, paste the error message directly into Claude or ChatGPT. Nine out of ten problems resolve in five minutes.

Final 15 minutes: Launch the first agent. Route 1: copy your goal statement into the system prompt field, run a sample input, look at the output. Route 2: say “Hello, agent” to Claude Code, then hand it your goal statement. Route 3: run the official tutorial’s Hello World sample as-is.

Whether it works in this hour: honestly, fifty-fifty. If it doesn’t run, that’s still normal progress. What matters is arriving at a state where you can see “how far did I get in 1 hour” and “what am I fixing in the next hour.” That’s the moment you crossed from user to builder.

What to do in the following weeks: once one agent runs, move to building a second. Only after building two in the same route do you really own it. If a week passes with nothing running, step down a route — Route 3 → 2 → 1. This isn’t failure. It’s correct tactical retreat.

For the next level — “10-person team all using the same agent,” “AI agents as a product to sell” — that’s a different conversation. The commercial direction, specifically how people without coding backgrounds build businesses around AI, is covered in From “using” to “selling”. After today’s roadmap gets you moving, that’s a natural next read.

Claude Jitsumujutsu (Uravation / Sato Suguru, releasing July 2026) looks like it’ll reinforce the direction for people at that stage. But run at least one agent before reading the book. Order matters.

Conclusion: 3 Routes Are Purpose-Specific. Start in 1 Hour.

AI agent building breaks into three routes. No-code Route 1. Low-code Route 2 with Claude Code. Full-code Route 3 using frameworks.

These aren’t a hierarchy. They’re purpose-specific paths you choose by reverse-engineering from your one-week goal. Not “Route 1 because I can’t code” — “Route 1 because I need one task automated.” Not “Route 3 because I’m an engineer” — “Route 3 because the goal is company-wide infrastructure.” That reframe is what prevents the first failure.

Four checks, cleanly. Goal clarity. Code experience. Budget. One-week target state. Run all four and the choice is unambiguous.

And today: open your phone notes, write one sentence about what you want to automate. Pick a route, set up a tool, launch your first agent. One hour.

People who do this today versus people who do it tomorrow will be in positions three times further apart six months from now. Not a threat — just the math. The fastest route to the “builder” side of AI agents isn’t reading about it, or attending events about it. It’s the one hour you have right now.

I still remember exactly how the day felt when my first agent ran. “There’s no going back from this.” I knew it immediately. I hope your version of that moment comes before today is over.