Build Your First AI Agent

Paragon AI MAG, Series 2, Episode 2 — "AI Agents: The AI That Does the Work"

OpenAI is shutting down Agent Builder — the drag-and-drop part of AgentKit for building agents inside ChatGPT — on 30 November 2026. Anyone who built an agent on it has weeks to rebuild somewhere else, or their agent simply stops working.

That is the whole lesson of this episode in one sentence: an agent you build on a tool you own survives. An agent built on a vendor's toy can be switched off overnight.

So this episode answers two questions: how do you build your first AI agent — with no code, for free, this week — and how do you build one that does not break, does not overspend, and does not embarrass you?

The 5-part recipe

A working first agent needs exactly five things:

1. One job, written as a clear outcome. "Watch this flight's price and tell me when it drops below 300 dollars" works. "Help me travel" does not.
2. Plain-English instructions that leave no room for interpretation. If you cannot write the job in five clear sentences, the agent cannot do it either.
3. One to three tools — email, calendar, web search. Every extra tool multiplies the ways it can fail.
4. Basic memory for the current conversation. A first agent needs almost no memory. Do not build a brain before you have a job.
5. A trigger with a strict approval step — the agent drafts, you approve before anything real happens. This one step is what separates a useful agent from a disaster. Episode 1 showed what happens without it: an agent deleted a whole production database during a code freeze.

The loop underneath it all takes fifteen seconds to say: see, think, act, check, repeat — until the goal is done or the agent needs your help.

The math that limits every agent

Keep the tool count low, because the math is brutal. If your agent takes five sequential steps and each step is 90% reliable, it succeeds only 59% of the time overall. That is 0.9 to the power of 5. Episode 1 showed the same math at twenty steps: 95% per step is only 36% overall. Long jobs fail because small errors stack up — so a first agent must have short jobs and few tools.

The builders, with real prices

You do not need to write a single line of code. The episode compares the visual builders with prices checked in October–November 2026:

• Zapier Agents is the easiest start for a non-technical person. It connects to more than 9,000 apps and gives 400 free activities a month; the Pro plan is $33.33 a month billed yearly. The warning: those 400 activities run out fast if your agent checks a database or an inbox every five minutes.
• Make gives a generous 1,000 operations free, and its Core plan is $9 a month for 10,000 credits. But Make counts every single step as one operation — a 12-step automation run 5,000 times a month consumes 60,000 credits. Adding one tiny step raises the bill immediately.
• n8n can be self-hosted free with unlimited executions (Community Edition, 400+ native nodes), or Cloud Starter at €20 a month for 2,500 executions. The key difference: n8n charges per full workflow run, not per step — a 20-step agent costs the same as a 2-step one.
• Custom GPTs inside ChatGPT are the fastest first step for drafting and research agents — but they live inside OpenAI's platform, which brings us back to the shutdown lesson.
• Agent-first tools — Relevance AI, Gumloop, and Lindy (which gives $50 in free credits for 7 days) — are for when the big three feel wrong, not as a first step.
• The code path exists for later: the OpenAI Agents SDK, version 0.24.0, needs Python 3.10 or newer. You do not need it for a first agent.

The build: a price-drop watcher in 6 steps

The episode builds one real agent, step by step:

1. Pick the job: a price-drop watcher for a flight or a product you actually want to buy. One job, one outcome, one number that tells you it worked.
2. Pick the builder: if you already live in Zapier, start there; if you want free and unlimited, n8n self-hosted; if you want a chat window, a custom GPT inside ChatGPT.
3. Write the instruction, word for word: "Check the price of this flight every morning at 8. If it drops below 300 dollars, email me the new price and the link. Never book anything. Never spend money."
4. Connect exactly two tools: a price check and an email notification. Two tools, not ten.
5. Test on five real examples before it touches anything live: five prices, five dates, five emails. Check each result in 30 seconds. If one fails, fix the instruction, not the model.
6. Set the guardrails: read-only first, a spending cap of zero, no payment card connected, no delete permission, a separate test account for anything that writes.

The cost math: a simple agent run costs a few cents in model calls. Zapier's free plan gives 400 agent activities a month, Make gives 1,000 operations a month free, n8n self-hosted is free with unlimited executions. A first agent costs zero dollars.

Why 40% of agent projects fail

This is the honest part. Gartner, in June 2025, predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 — because of escalating costs, unclear business value, and weak risk controls. Not because agents are fake.

The adoption numbers behind that prediction are stark. Gartner's 2026 CIO survey shows only 17% of organizations have actually deployed agents, even though more than 60% expect to within two years. Capgemini surveyed 1,500 executives at companies over $1 billion in revenue: only 2% have fully scaled AI agents, and 14% have started implementation. Gartner estimates that only about 130 of the thousands of companies calling themselves "agentic AI" vendors are real — the rest is agent washing, old automation repackaged with a trendy name.

Trust is falling, too. Trust in fully autonomous agents fell from 43% to 27% among executives in one year. A SailPoint survey of 353 IT professionals found that 80% said their AI agents had already done something unintended — and 23% said their agents were tricked into revealing sensitive access credentials. Yet only 44% have policies in place to secure them.

The enterprise side is moving anyway: KPMG's Q3 2026 pulse found 62% of large US enterprises report agents in deployment, with 25% still piloting. The gap is between pilot and production — and that gap is where the money is.

The "sell agents" trap

A video with 3.91 million views promises "build and sell AI agents" as a business. The honest math: the build is the easy 10%. Support, reliability and trust are the job. Agents break when websites change, clients blame you when the agent misbooks or overspends, and maintenance is the business, not the build. Before you deploy anything critical, ask: who gets the alert when this breaks on a Tuesday at 6pm?

Your first agent this week

Three starter agents to copy:

1. The price-drop watcher from the build above.
2. Inbox triage — the agent drafts replies, you approve before sending.
3. A daily news summarizer sent to your phone each morning.

And the strict rule before any of them touches anything live — the 5-minute test. Run your agent on five real examples. Verify each result manually in 30 seconds. It must achieve five straight passes before it touches anything live. Five straight passes. That is a non-negotiable rule.

What not to automate first: anything that spends money, deletes data, or sends messages to customers without your approval. And run the 3-question test every time: can I check the result in 30 seconds? Can I undo it? Is the worst mistake cheap? If any answer is no, do not delegate it yet.

Paragon's verdict

🟢 CLEARED — build it. A small first agent with a leash is worth building this week: zero dollars to start, free tools at this scale, and the approval step makes it safe.

🔴 SKEPTICAL — the "sell agents and get rich" promise. The build is the easy part; support, reliability and trust are the business.

Next in the series: Episode 3 — "Apps Are Dying? The Microsoft CEO's Prediction, Explained Simply."

Originally published on Substack: https://paragonaimag.substack.com/p/build-your-first-ai-agent

This article is the written version of Paragon AI MAG, Series 2, Episode 2. Watch the video on the Paragon AI MAG YouTube channel — https://youtu.be/yVZ06ecrDtc — and listen to the full 47-minute podcast conversation on this topic.