What prompt designs an AI agent before you build one?
Paste this into Claude, ChatGPT, Gemini or Whizi and fill the brackets. It forces the one job, the tools and the stop rules onto a page before you open any builder.

Design an AI agent for me. The one job is: [one sentence, with a result I can check]. The inputs it gets are: [what arrives, such as an email, a ticket or a row in a sheet]. The tools it may use are: [apps or files], and for each one say read only or can write. List every action that needs a human to approve it first. Write its instructions as 8 rules that each have a pass or fail test. Then write 10 test inputs, including 3 it must refuse, and name the step limit you would set.
The output's a spec you can hand to any platform in this guide. It is also the part most first agents skip, and the part that decides whether the agent works.
What is OpenAI Agent Builder, and is it still available?
Agent Builder is OpenAI's visual canvas for multistep agent workflows, and it is being shut down. OpenAI's Agent Builder docs say OpenAI is deprecating it and that the product is scheduled to shut down on November 30, 2026, with existing users able to keep using it in the meantime. The deprecations page lists the notice date as June 3, 2026 and recommends the Agents SDK or ChatGPT Workspace Agents. Both pages were read on November 22, 2026, eight days before the date.
AgentKit is the bundle Agent Builder belonged to. OpenAI's AgentKit announcement names three parts: Agent Builder, Connector Registry for admins, and ChatKit for embedding a chat agent in your product. ChatKit remains available, according to the docs. The announcement's June 3, 2026 update says Agent Builder and Evals are being wound down and will not be available on the OpenAI platform from November 30, 2026.
On cost, the announcement says the tools are included with standard API model pricing. So there was no per seat or license fee, and you paid for the model calls your agent made. That is why a search for the openai agent builder pricing has no price table to find. For what the models cost, see OpenAI API pricing, and for the key every API call needs, how to get an OpenAI API key.
Many top results for this query still describe Agent Builder as a current product. One guide in the top five is dated October 9, 2025, and a vendor blog from November 2025 calls Agent Builder a public beta and lists "coming soon" features. Neither mentions the shutdown, so read their steps as history.
How do you build a first agent in OpenAI Agent Builder?
If your account can still open it, the build runs in three stages, per OpenAI's docs: design, publish, deploy. Treat this as a record of how it worked, because it stops on November 30, 2026.
- Start from a template or a blank canvas. The docs use a homework helper template that rewrites a question, routes it to a specialist agent and returns an answer.
- Add nodes for each step and set typed inputs and outputs. Connections between nodes become typed edges.
- Preview the run with live data, attach sample files and watch each node execute.
- Run trace graders from the top navigation to judge how the workflow did.
- Publish. This creates a workflow object with an ID and a version, and each publish makes a new major version you can still call by number.
- Click Code in the top navigation. Pass the workflow ID into a ChatKit integration, which the docs recommend if you are unsure, or copy the workflow code and use it anywhere through the Agents SDK.
The exported code is the way out. A workflow you built there can live on as code, which is the route OpenAI's own announcement recommends for workflows that should continue. Saving a copy of the exported code before November 30 is the one urgent step on this page.
How do you build an AI agent from scratch?
You build an AI agent by giving a model one job, a short list of tools and rules you can test, then running it on 20 inputs before it touches real data. The six steps in the how to list above hold on every platform.
The code route is the shortest to explain. OpenAI's Agents SDK docs describe a Python package installed with pip install openai-agents. You create an Agent with a name and instructions, then call Runner.run_sync with an input and read final_output. The page says to set the OPENAI_API_KEY variable first and lists the other pieces as tools, handoffs between agents and guardrails that check inputs and outputs. It states no price, so the cost is the model calls on your own API account.
A tool is how an agent acts, and MCP is the common way to plug one in. What is MCP explains it, and what is an AI agent covers the difference between an agent and a chatbot.
Which is better, OpenAI Agent Builder, n8n or Zapier Agents?
For a new build, n8n or the Agents SDK, because Agent Builder will not exist after November 30. The table reads each vendor's own page on November 22, 2026, and a cell says not stated where the page was silent.
| Tool | Status | Free use | Paid price | You need | Best at |
|---|---|---|---|---|---|
| OpenAI Agent Builder | Deprecated, shuts down Nov 30, 2026 | No separate fee | Standard API model pricing | An OpenAI API account | Nothing new, export the code |
| OpenAI Agents SDK | Python package | Not stated | Not stated, model calls bill to your key | Python and an API key | Agents you own as code |
| ChatGPT Workspace Agents | Research preview | Not stated | Credit based pricing began May 6, 2026 | Business, Enterprise, Edu or Teachers plan | Team agents that live in ChatGPT and Slack |
| n8n | Live | No free plan, trial of 1,000 executions, self hosted Community Edition | Starter $20 a month billed annually, Pro $50 | A hosted plan or your own server | Visual workflows you can host yourself |
| Zapier Agents | Live | Free, 400 activities a month | Agents Pro about $33.33 a month billed annually, 1,500 activities | A Zapier account | Agents across 9,000+ connected apps |
| Whizi | Chat app, not an agent builder | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card | Pro $29.99 a month | An account | Drafting the spec, tests and files |
Whizi loses on the main question. It doesn't build or run agents, so it can't replace any row above it. n8n wins if you want self hosting, since its pricing page links a self hosted Community Edition on GitHub and shows no price for it. Zapier wins if your apps are already connected there, and its pricing page counts agent activities, not tasks, with a cap of 10 activities a run on Free and 40 on paid plans.
OpenAI points to Workspace Agents for uses better suited to plain language prompting. Its launch page calls them an evolution of GPTs that run in the cloud and can be reached in ChatGPT or Slack. It lists the Business, Enterprise, Edu and Teachers plans, so it is not an option on a personal plan.
How do you build an n8n AI agent?
You drop an AI Agent node on the canvas, connect a chat model, and connect at least one tool. n8n's docs say you must connect at least one tool sub node, and that the agent chooses which tool to call for the task. Since version 1.82.0 every AI Agent node works as a Tools Agent, and the old agent type setting is deprecated.
The pricing page lists Agents, MCP and AI nodes on the hosted Starter and Pro plans, at $20 and $50 a month billed annually for 2,500 and 10,000 workflow executions. There is no free plan. It offers a trial with no card and a self hosted Community Edition on GitHub. The step list from the section above applies unchanged: one job, listed tools, testable rules, then 20 test inputs.
The tradeoff against the SDK is control against effort. n8n gives you a canvas and ready connectors. The SDK gives you code you can version and test like any program, which matters when an agent handles money or customer data.
What does a good answer look like for a first agent?
These three examples were written for this page, not run through a named model. Each shows the prompt and what a good reply must contain, so you can tell a usable spec from a vague one.
- Support triage. Prompt: "Use the design template to specify an agent that reads each new support email, tags it as billing, bug or other, and drafts a reply for me to approve. It must never send anything." A good answer lists read only email access, one write action (save a draft), a rule that refunds are always routed to a human, and 10 test emails including a message with no question in it.
- Weekly research brief. Prompt: "Specify an agent that searches the web every Monday for news about [topic], keeps only items from the last 7 days, and writes a one page brief with a link for each claim." A good answer sets a source rule, a date filter, a step limit, and a test where the search finds nothing and the agent must say so instead of filling the page.
- Invoice chaser. Prompt: "Specify an agent that checks a sheet of unpaid invoices and drafts a polite reminder for any more than 14 days late." A good answer makes the sheet read only, writes drafts to a separate file, never emails by itself, and tests a row with a missing date, which the agent must flag.
What prompts help you design, test and migrate an agent?
Twelve prompts, grouped by the stage you are at. Fill the brackets and paste them into any chat model. The first one is the template above.
Design
- Spec an agent: the template in the first section.
- Cut the scope: "Here is my agent idea: [idea]. List what to remove so it does one job, and say what each removed piece should become, a second agent or a human step."
- List the tools: "For this agent spec, list every tool it needs, whether each is read or write, and which actions need a human to approve them: [paste spec]."
Test
- Write test cases: "Write 20 test inputs for this agent: 10 normal, 5 messy and 5 it must refuse. Give the expected result for each: [paste spec]."
- Review a run: "Here is the step log of one run: [paste]. Find the first step where the agent went wrong, say why in one sentence, and propose one change to the instructions."
- Make the test sheet: "Put these test cases in a table with the columns input, expected result, actual result and pass or fail, ready to paste into Excel: [paste cases]."
Rules and cost
- Make rules testable: "Rewrite these agent instructions so each rule has a pass or fail test, and mark any rule that cannot be tested: [paste]."
- Set approval gates: "Read this workflow and list each step that sends, pays, deletes or shares something, then tell me where a person must approve: [paste]."
- Count the calls: "Estimate how many model calls one run of this agent makes and write the formula, so I can multiply by the price on my vendor's page. Do not invent prices: [paste spec]."
Move and choose
- Read exported code: "Here is the code exported from my Agent Builder workflow: [paste]. Explain each part in plain words and list what to change to run it with the Agents SDK."
- Pick a platform: "I need an agent that [job]. I [can or cannot] write code and my team uses [apps]. Compare the Agents SDK, n8n and Zapier Agents using only the pricing text I paste: [paste]."
- Write the handover: "Write a one page note for my team explaining what this agent does, what it may never do, who approves its actions and how to switch it off: [paste spec]."
What mistakes ruin a first AI agent?
Five mistakes cause most failed first agents, and each has a fix.
- Building before you have a spec. The fix: run the design prompt first and build only from its output.
- Giving the agent write access it does not need. The fix: start read only and add one write action at a time, with approval.
- Testing only the happy path. The fix: include inputs that should be refused, because that is where agents do damage.
- No step cap on a loop. The fix: set a maximum number of steps, and for Zapier note the cap of 10 activities a run on Free.
- Copying a tutorial about a product that is leaving. The fix: check the vendor's deprecations page before you learn a tool, as this guide did for Agent Builder.
Where does Whizi fit when you build an agent?
Whizi is one app instead of a pile of single purpose tools, and for agents it is the place you plan, test and document, not the place you run one. It doesn't build or host agents, so use the platforms above for that. What it saves you is the setup around them: no API key, no billing account and no rate limit tier to manage just to draft a spec or test a prompt.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. On Pro and Powerhouse, side by side Compare puts one prompt to two models, which is a quick way to check a spec or compare two sets of agent instructions. See Whizi pricing for each plan.
Excel and Word files made in chat have no plan gate, so prompt 6 can hand you the test sheet and prompt 12 the handover note as files. You can also upload a PDF, Word or Excel file for the model to read, such as a vendor policy or a log. Web search is on by default per chat, which helps when you check a vendor page. If you want to compare chat assistants first, see our best AI agents guide.
What can an AI agent not do reliably?
It can't be trusted with an irreversible action unaided. An agent picks its next step from likely text, so it can pick a wrong tool, loop, or act on a message that says to ignore its rules. That's why Agent Builder shipped guardrail and approval pieces, and why this guide keeps approvals on for anything that sends, pays or deletes.
It also can't tell you a vendor's price or date from memory. Every price in the table came from a vendor page on November 22, 2026, and plans change. Check the page before you buy. If an agent will touch customer data, money or health records, a qualified person has to review it first, and the vendor's own terms apply.
- Save the exported code of any Agent Builder workflow before November 30, 2026
- Write the one job and its checkable result before opening a builder
- Start every tool read only and add write access one action at a time
- Run 10 normal, 5 messy and 5 refusal inputs before real data
- Cap the steps and approve anything that sends, pays or deletes
Frequently asked questions
Is OpenAI Agent Builder free?
It had no separate fee, but you paid for model use. OpenAI's AgentKit announcement says the tools are included with standard API model pricing, so each run is billed at API rates. It is also shutting down on November 30, 2026, so it is not a tool to start on now.
What replaces OpenAI Agent Builder?
OpenAI recommends the Agents SDK for workflows that continue as code and Workspace Agents in ChatGPT for plain language use. Outside OpenAI, n8n and Zapier Agents both build agents on a canvas, as of November 2026. Export your workflow code before the shutdown.
Is OpenAI AgentKit the same as Agent Builder?
No. AgentKit is the bundle and Agent Builder is one part of it, next to Connector Registry and ChatKit. Per OpenAI, only Agent Builder and Evals are being wound down on November 30, 2026, and ChatKit remains available.
Do you need to code to build an AI agent?
No. n8n and Zapier Agents work on a canvas or in plain language, and Agent Builder was a drag and drop canvas. Code gives more control: the OpenAI Agents SDK is a Python package that needs an API key. Either way, test with 20 inputs first.