How to build an AI agent, with or without code
A practical guide for non-developers and developers alike: what an agent really is, a no-code path and a code path, the tools with free plans as of October 2026, and the limits most guides skip.
How do you build an AI agent?
Pick one narrow job, then choose a no-code builder such as Zapier Agents or n8n, or a code framework such as the OpenAI Agents SDK. Give the agent clear instructions, connect only the tools it needs, require your approval before it sends or pays, and test it on real examples before letting it run.
- Search demand
- About 5,900 US Google searches a month for “how to build an ai agent” (Ahrefs, October 2026)
- What an agent is
- An AI model with instructions and tools that decides its own next step until a job is done
- No-code, free to start
- Zapier’s free plan includes Agents with 400 activities a month; n8n’s Community Edition is free to self-host (October 2026)
- Code, free to start
- OpenAI Agents SDK, LangGraph and CrewAI are MIT-licensed; you still pay for the AI model you call
- Tool to avoid starting on
- OpenAI’s Agent Builder, which OpenAI says is scheduled to shut down on November 30, 2026
What is an AI agent, and what do you need to build one?
An AI agent is an AI model that is given a goal, instructions and a set of tools, and then decides for itself which step to take next until the job is done. A chatbot answers a message. An agent reads your inbox, looks up the customer, drafts a reply and asks you to approve it.
Anthropic draws a useful line in its guide to building effective agents: workflows follow steps you define in advance, while agents direct their own process and tool use. Its advice is to find the simplest solution that works, because agents trade speed and cost for flexibility. Many “agents” people want are really agentic workflows, and that is a good thing: they are cheaper and more predictable.
- A model: the AI that reads, reasons and writes, such as GPT, Claude or Gemini.
- Instructions: the job, the rules, the tone, and what to do when unsure.
- Tools: the actions it can take, such as search the web, read a sheet, send an email or create a ticket.
- Knowledge: the documents or data it should use instead of guessing.
- Limits: what needs your approval, how many steps it can take and how much it can spend.
If you are still unsure where agents end and “agentic AI” begins, read AI agents vs agentic AI first.
Should you build an AI agent with or without code?
Build without code if the agent works inside apps you already use, like Gmail, Slack, a CRM or a spreadsheet. Write code if the agent is part of a product you sell, needs custom logic, or has to run at a volume where per-task pricing gets expensive. This table helps you decide.
| Your situation | Better path | Why |
|---|---|---|
| You don’t code and want an agent for your own work | No-code builder | Visual setup, ready-made connections to common apps, nothing to host |
| The agent connects to apps a builder already supports | No-code builder | Connecting apps is most of the work, and the builder has done it |
| You need custom logic, your own database or unusual APIs | Code | You control every step, and nothing is limited to a builder’s menu |
| You want to sell the agent to customers | Code, or an app builder that includes sign-in and payments | Customers need accounts, billing and support, which automation tools don’t provide |
| High volume, thousands of runs a day | Code or self-hosted n8n | Per-task and per-credit pricing adds up; self-hosting moves the cost to your server and model bill |
| Sensitive data such as health or finance | Code, with a developer review | You need to know exactly where data goes and who can see it |
What tools can you use to create an AI agent?
You can create an AI agent with a no-code builder, an automation platform with AI steps, or a code framework. Here are real options and their free starting points, checked on each company’s own pricing page or docs on October 24, 2026.
| Tool | Type | Free option | Good for |
|---|---|---|---|
| Zapier Agents | No-code | Included on the free plan: 400 activities a month, up to 10 per agent run | Agents that act across many apps; Zapier says 9,000+ |
| n8n | No-code / low-code | Self-hosted Community Edition; cloud Starter is €20 a month billed yearly, with 2,500 executions | Multi-step agents with tools, memory and MCP servers |
| Lindy | No-code | $50 in credits for 7 days, one time; Team plan from $49.99 a month | Assistants for email, meetings and scheduling |
| OpenAI Agents SDK | Code (Python) | Free, MIT license; you pay for model use | Agents with tools, handoffs between agents and guardrails |
| LangGraph | Code | Free, MIT license; you pay for model use | Long-running agents that keep state between steps |
| CrewAI | Code (Python) | Free, MIT license; you pay for model use | Teams of agents with different roles |
| Claude Agent SDK | Code (Python, TypeScript) | Library is free to use under Anthropic’s Commercial Terms; you pay for API use | Agents that read files, run commands and use MCP, with permission controls |
How to build an AI agent, step by step
- 1
Pick one narrow job
Write the job in one sentence with a clear finish line: “Reply to every new lead within five minutes and book a call,” not “handle sales.” One job is easier to test, cheaper to run and easier to trust.
- 2
Do the job by hand three times
Write down every step, every app you open and every decision you make. This becomes the agent’s instructions and tool list, and it shows you whether a fixed workflow would be enough.
- 3
Choose your path and tool
Use the decision table above. If you don’t code, start with Zapier Agents or n8n. If you do, start with one SDK and one model, not a stack of frameworks.
- 4
Write the instructions
Describe the goal, the steps, the tone, what “done” looks like and what to do when unsure. Add two or three examples of good outputs. Be specific about what the agent must never do.
- 5
Connect only the tools it needs
Give read access first and write access only where the job needs it. An agent that only drafts emails can’t send a wrong one.
- 6
Add approvals and limits
Require your OK before anything is sent, charged, published or deleted. Cap the number of steps per run and set a daily spending limit with your model provider.
- 7
Test on real examples
Run it on 20 real cases from your history, including messy ones. Note each wrong answer, fix the instructions, and run the same 20 again until the results stop surprising you.
- 8
Launch small and review the logs
Turn it on for a slice of the work, read what it did every day for the first week, and loosen approvals only for the actions it gets right every time.
How do you build an AI agent without code?
To build an AI agent without code, you use a visual builder: you describe the job in plain language, pick the apps it can use from a menu, and test it in a chat window. Here is what that looks like for a common first agent, a lead follow-up assistant.
- Trigger: a new form submission or a new email with “quote” in the subject.
- Instructions: “Read the lead’s message, look them up in the CRM, and draft a friendly reply that answers their question and offers two call times. Never quote prices that aren’t in the price sheet.”
- Tools: read the CRM, read the price sheet, read your calendar, create an email draft.
- Approval: drafts go to your inbox; you press send.
- Test: forward ten old leads and compare the drafts with what you actually sent.
The limit of no-code is the edges: if the builder doesn’t support an app or an action, you wait or work around it. And per-task pricing that is cheap for ten runs a day can be expensive for ten thousand.
How do you make an AI agent with code?
To make an AI agent with code, you call a model in a loop: send it the goal and the list of tools, run whichever tool it asks for, send back the result, and repeat until it says the job is done or hits your step limit. Frameworks write that loop for you and add the parts that are easy to get wrong.
- The OpenAI Agents SDK describes agents as LLMs “equipped with instructions and tools,” with handoffs between agents and guardrails that check inputs and outputs.
- LangGraph is a lower-level framework for agents that keep state and run over long periods.
- The Claude Agent SDK gives you the same tools, agent loop and permission controls that power Claude Code, in Python or TypeScript.
- Tools are usually functions you write, or servers that speak MCP, the open standard for connecting AI apps to outside data and tools.
Two habits save most code-path agents: log every step the agent takes, so you can see why it did something, and write a small test set of real cases you rerun after every change to instructions or model version.
What are the limits of AI agents?
AI agents are limited by cost, reliability and trust. Each step is a paid model call, mistakes compound over many steps, and an agent with access to your email or payments can do real damage when it gets something wrong. Anthropic warns that agents bring higher costs and the potential for compounding errors.
| Limit | What happens | What to do |
|---|---|---|
| Cost | One task can take many model calls, and a loop that never finishes keeps billing | Cap steps per run, set a spending limit, and check the cost per task before launch |
| Reliability | The same input can give different results, and small errors stack up across steps | Keep jobs narrow, test on real cases, and use a fixed workflow where steps never change |
| Approvals | An agent can send, pay or delete faster than you can notice | Require your OK for anything external or irreversible |
| Data and access | Broad access means a mistake or a malicious input can reach everything | Give the fewest permissions possible and separate accounts per agent |
| Platform change | Prices, limits and whole products change or close | Keep instructions and test cases in your own files so you can move |
What if you want to sell your AI agent?
If you want customers to use your agent, you need more than the agent: sign-in so each person has their own data, payments, usage limits so one customer can’t run up your bill, support and analytics. That is an app, not an automation.
You can build that yourself, or use an AI app builder that includes the business parts. Our guide on how to build an AI app walks through pricing, launch and running it.
Build an AI agent your customers can pay for
Starty is our product. Describe the agent in one sentence, and Starty builds it inside an app with sign-in, payments, emails, a support agent, legal pages and analytics, all in test mode until you review.
- An AI agent builder with a goal, triggers (when asked, from a photo, on a schedule), knowledge from your text and files, tools, memory per person and a daily use limit per person
- Starty runs and meters the AI per use, so you see the AI cost per paying user
- Starty Payments built in for subscriptions and packs, with no Starty fee on your revenue
- Nothing is sent, charged, refunded or published without your OK
Related guides
Where every fact comes from
Official pricing pages, docs, announcements and named studies, checked on October 24, 2026. Where a company doesn’t list something, we say so rather than guess.
Questions people ask
Start on a free plan: as of October 2026, Zapier’s free plan includes Agents with 400 activities a month, and n8n’s Community Edition is free to self-host. Code frameworks like the OpenAI Agents SDK, LangGraph and CrewAI are free and open source, but you pay for the AI model calls.
Yes. No-code builders let you describe the job in plain language, pick the apps it can use and test it in a chat window. They work best when the agent lives inside apps the builder already supports, such as email, calendars, spreadsheets and CRMs.
A simple no-code agent for one narrow job can be working in an afternoon. Making it reliable takes longer: testing on real cases, fixing instructions and watching its first week of work. Agents with custom code, sensitive data or paying customers take much more.
You pay for the builder or hosting, plus the AI model calls the agent makes. An agent can make many calls per task, so cost depends on the job, the model and how many steps it takes. Cap steps per run and set a spending limit before you go live.
A chatbot answers messages in a conversation. An agent takes actions toward a goal: it can look things up, use other apps and decide its next step. Many products mix both, with a chat window on top of an agent that can act.
There is no single best one. The OpenAI Agents SDK is light and quick to learn, LangGraph suits long-running agents that keep state, CrewAI suits teams of agents with roles, and the Claude Agent SDK suits agents that work with files and commands. Pick one and learn it well.
Enter your email to join Starty’s private beta. We’ll email you your access, and you can start building for free.
No. You describe what you want, or fork an idea, and Starty builds it. You review and approve; you never have to open the code.
No. Starty takes no fee on your revenue. Payouts go to your bank account, and Starty earns from plans and credits.