How to Create an AI Agent (Plain English, No Code)
Garrett Pierson
To create an AI agent, write down the job you want done, pick a no-code builder like Zapier or Make, connect the apps the agent needs to touch, set a limit on what it’s allowed to do without asking you, then test it against work you already finished. Six steps, and a free plan is enough to find out whether it works.
The word “agent” makes this sound harder than it is. You’re hiring, and the hard part of hiring has never been the paperwork.
What you’re actually building
An agent is software you give a goal instead of a script. You tell a chatbot what to say; you tell an agent what you want to be true when it’s finished, and it picks the steps.
The dividing line is whether it needs you standing between the steps. A form that emails you when someone fills it out is automation. Something that reads the form, checks the person against your CRM, writes a reply in your voice, and books the call is an agent. We break down the best AI agents for business if you’d rather buy one off the shelf than build it.
Before you start
Two things, and neither is technical:
- A job that repeats. Agents earn their keep on work that happens weekly or daily. A one-time task is faster to just do.
- A way to tell if it worked. “Fewer missed leads” is measurable. “Better organized” is not.
Budget $0 to $40 a month. Every tool below has a free plan or a trial, and you pay for model usage on top of whatever the platform charges.
Step 1 — Write the job description
Before you open any tool, write four lines in a plain text file:
- The trigger. What starts the job? A new email, a filled-in form, 8am every Monday.
- The goal. What’s true when it’s finished?
- The tools. What does it need to see or touch? Gmail, your CRM, a spreadsheet, the web.
- The line it must not cross. What requires your approval?
This file is the whole build. Most agent tools ask you to paste something like it into a box, and a vague answer here produces a vague agent that does almost-right things forever.
Step 2 — Pick where the agent will live
Pick one. All five let a non-programmer get an agent running. We checked every price below on the vendor’s own pricing page in September 2026:
| Tool | Code needed? | Free tier | Paid from |
|---|---|---|---|
| Zapier | No | 100 tasks/mo | $19.99/mo annual, $29.99 month-to-month (750 tasks) |
| Make | No | 1,000 credits/mo | $12/mo Core (10,000 credits) |
| Lindy | No | No free plan; 7-day trial for teammates joining via Slack | $29.99/user/mo Plus (3,000 credits) |
| Gumloop | No | 14-day trial, no free plan | $37/mo Pro (20,000 credits) |
| n8n | Some, for anything advanced | Free, self-hosted | €20/mo Starter on annual billing (2,500 executions) |
Start with Zapier if the job spans apps you already use, or Make if you want a cheap paid plan and a visual canvas. Choose n8n when you’d rather run the whole thing on your own server.
One billing trap worth knowing before you pick: Zapier now runs agents as an AI step inside an ordinary Zap, and that step burns tasks at 1×, 3×, or 5× depending on which model you choose. Make counts credits, n8n counts workflow executions. The same job costs a different amount on each meter, so watch your first month’s usage before you scale anything up.
One to avoid right now: OpenAI’s Agent Builder. OpenAI announced its deprecation on June 3, 2026, and the shutdown lands November 30, 2026. The official migration path is the Agents SDK or ChatGPT Workspace Agents. OpenAI is winding down its Evals platform on the same date, while ChatKit stays available, so parts of AgentKit are going and parts are staying. Don’t start your first agent on the parts that are leaving.
Step 3 — Connect its tools, and only the ones it needs
Connecting an agent to your tools is what makes it useful, and it’s the part that decides whether this is safe.
Give it the narrowest access that lets it finish the job. A dedicated email address rather than your inbox. Read-only on the CRM if it only needs to look things up. A separate API key you can revoke on a bad day without breaking anything else.
If you want an agent reaching software that isn’t on your tool’s integration list, that’s what MCP servers are for — a standard plug shape that lets any AI tool talk to any outside service.
Step 4 — Set the guardrails before you turn it on
Write the rules into the agent’s instructions, in plain English, before its first run:
- What it must never do without you. Spend money, contact a customer, delete anything.
- What it does when it’s unsure. “Flag it and stop” beats a confident guess every time.
- How much it can do per run. A cap of ten emails means a bad Monday costs you ten emails.
Zapier, Gumloop, n8n, and Make all support a human approval step. Turn it on for the first month even when you think you won’t need it.
Step 5 — Test it against last week’s real work
Run the agent on last week’s leads, last week’s invoices, last week’s support tickets. You already know the right answers, so grading it takes minutes.
Watch for the failure that matters: the near-miss. An agent that crashes tells you it broke. An agent that files the invoice under the wrong client, politely and on schedule, will do that for months before you notice. Read ten full outputs end to end before you trust the eleventh.
Step 6 — Put it on a trigger and check it weekly
Move it off manual runs and onto the trigger you wrote in Step 1. Then put a calendar reminder to review its output every week for a month.
Agents drift. Your CRM changes a field name, a vendor updates an API, a client starts sending PDFs instead of spreadsheets, and the agent keeps running against a world that moved. The weekly check is how you catch that in week one instead of quarter three.
If you’d rather build an AI agent with code
If you searched for “how to build an AI agent,” you probably landed on developer tutorials. That path is open to you now in a way it wasn’t three years ago, because you can generate most of the code by describing it.
Anthropic’s Claude Agent SDK (renamed from the Claude Code SDK in September 2025) and OpenAI’s Agents SDK both give you the agent loop, tool calling, and memory as building blocks in Python or TypeScript. You get control the no-code tools won’t give you, and you take on hosting, error handling, and everything that breaks at 2am. If that’s the direction you want, building an app with AI walks the same workflow, and the best AI coding agents covers what writes the code for you.
For a first agent, the no-code path is the shorter road.
What trips people up
- Building the five-agent system first. Start with one agent on one boring repeating task and give it a month. A five-agent system is slow to debug and expensive while you’re debugging it.
- Skipping the job description. A prompt like “manage my inbox” produces an agent that does something inbox-shaped and nothing you asked for.
- Giving it your whole account. Full admin access to a CRM is how a small mistake becomes a big one.
- Trusting a demo run. It worked once on a clean example. Grade it on ten messy real ones.
- Never checking the bill. Credits and tasks meter differently on every platform. Watch the first month’s usage before you scale the agent up.
The bottom line
Creating an AI agent takes an afternoon and, on a free plan, nothing: write the job description, pick a no-code builder, connect the narrowest set of tools that works, cap what it can do alone, and grade it against work you’ve already done. The tooling is no longer the hard part.
What an agent can’t do is decide what your business should be building. That part is yours, and it’s what Software Secrets 2.0 covers, start to finish, for free. If you’re thinking about selling agent builds to other companies, starting an AI automation agency covers what that business actually pays.