How to Create an AI Agent with ChatGPT Agent Builder (2026 Guide)

Key Takeaways

• "ChatGPT agent builder" can mean three different products: Agent Builder (the visual AgentKit canvas), Custom GPTs (the no-code builder inside ChatGPT), or ChatGPT Agent Mode. Confirm which one you need before you start.
• OpenAI announced on June 3, 2026 that Agent Builder is being wound down, with a shutdown date of November 30, 2026. Existing workflows keep running during the transition, but new long-term projects should not start here.
• Building in Agent Builder is a visual, node-based process: configure a Start node, add Guardrails, set up the Agent node, connect tools and MCP servers, add logic, then test and publish through ChatKit or the Responses API.
• OpenAI's recommended replacements are the Agents SDK for code-first workflows and Workspace Agents in ChatGPT for natural-language agents. You can export a workflow to Agents SDK code, but the export does not convert the graph or guarantee identical behavior.
• For a customer support agent, rebuilding your own agent every time a vendor retires a tool is rarely worth it. A purpose-built B2B support platform delivers the same outcome without the maintenance churn.

OpenAI announced on June 3, 2026 that it is winding down Agent Builder (and the Evals product), with a scheduled shutdown date of November 30, 2026, confirmed on OpenAI's deprecations page.

Existing workflows continue to run during the transition window, and ChatKit's advanced integration stays available. What goes away is the canvas itself and the OpenAI-hosted backend it powered.

Here is how to use the rest of this article. If you are maintaining an existing Agent Builder workflow, or you want to understand the model before you migrate, read the step-by-step build below.

If you are starting from scratch today, skip ahead to what to use now that Agent Builder is going away and start there instead.

Either way, build with the migration path in mind. Nothing below is wasted knowledge, but the destination has changed.

Which "ChatGPT Agent Builder" Do You Actually Mean?

Three different OpenAI products get called the "ChatGPT agent builder," and they are not interchangeable. Pick the wrong one and you will follow steps that do not match your screen. Here is how they line up.

If you want a workflow engine with branching logic and tool connections, you mean Agent Builder, and the steps below are for you.

If you just want a configured assistant trained on your documents, you actually want a Custom GPT, which we cover near the end.

If you want ChatGPT to run a task for you right now, that is Agent Mode, and you do not need a builder at all.

What Is OpenAI Agent Builder (AgentKit)?

Agent Builder is OpenAI's visual canvas for building multi-step agent workflows by dragging and connecting nodes.

It launched as part of AgentKit at DevDay 2025, and it uses a small set of node types (Start, Agent, Guardrail, Logic, Tools, and Data) that you wire together, preview with live data, then publish and embed with ChatKit.

The appeal was the blank canvas. Instead of writing an agent loop and tool-execution code yourself, you assembled the flow visually and let OpenAI host the backend.

That is also the part being retired, which is why the migration section matters as much as the build steps.

How to Create an AI Agent with ChatGPT Agent Builder (Step by Step)

The flow below still works today and is worth knowing if you are maintaining an existing workflow or learning the model before you move it.

Treat it as "how it works right now," and weigh the migration section before starting anything new on it.

Step 1: Access Agent Builder and Start a Workflow

Go to platform.openai.com/agent-builder and sign in with your OpenAI account. Open your account settings and add billing details, which you need to enable Preview mode.

The interface has three tabs. Workflows holds your published agents, Drafts holds works in progress, and Templates offers prebuilt starting points.

Click Create New Workflow to begin from a blank canvas, or start from a template such as Customer service or Data enrichment.

Step 2: Configure the Start Node

Every workflow opens with a Start node that defines the inputs your agent receives. Click it to set input variables (user-provided values like "destination" or "travel_dates") and state variables that persist across the conversation (like "user_preferences" or "budget_range").

State variables are what let the agent hold context across turns. Conversation history sits in a short-term buffer, while structured long-term memory retains the details you want the agent to remember.

Step 3: Add Guardrails

Drag a Guardrail node onto the canvas and connect it to the Start node. This is where you turn on the safety checks that keep the agent compliant.

• PII detection and redaction. Automatically catches and masks sensitive data like names, phone numbers, and account IDs before they move through the workflow.
• Content moderation. Blocks harmful or inappropriate inputs from reaching the agent.
• Jailbreak detection. Flags prompt-injection attempts that try to override your instructions.
• Hallucination checks. Verifies responses to reduce confident-but-wrong answers.

A healthcare scheduler, for example, would enable PII redaction to protect patient details and moderation to keep interactions professional.

Step 4: Set Up the Agent Node

The Agent node is the core of the workflow. Drag it onto the canvas and connect it to your Guardrail node, then configure how it thinks.

• Name. Give it something descriptive, like "Healthcare Appointment Assistant."
• Instructions. Spell out the role, tone, and limits. For example: "You are a medical appointment scheduler. Check availability, suggest slots, confirm bookings, and always ask for patient ID. Be empathetic and professional."
• Model. Choose based on the job. Heavier models handle complex reasoning, while a smaller model like GPT-5-mini is faster and cheaper. Pick the simplest model that does the job reliably.
• Reasoning effort. Set lower for speed, higher for harder tasks.
• Output format. Choose text for conversation, JSON for structured data, or a widget for interactive replies.

Step 5: Connect Tools and MCP Servers

To give the agent real capabilities, connect tools through the Connector Registry. File Search enables retrieval over uploaded documents, so you can point the agent at product catalogs, FAQs, or policy manuals.

MCP servers connect the agent to external apps like Gmail, Google Drive, Slack, Notion, or your CRM. You authenticate each service once, and the agent can then act through it.

A marketing agent might connect Gmail for campaigns, Google Sheets for tracking, and Slack for notifications.

Step 6: Add Logic and Branching

Use logic nodes to handle different paths through the workflow. If/Else nodes route based on conditions, While nodes repeat steps, and User Approval nodes pause for a human to confirm before the agent acts.

For example, after processing an order you might branch on value. If the order is over $500, route it to a User Approval node for manager sign-off. If not, proceed straight to confirmation. Chaining smaller, specialized agents this way also makes the workflow easier to debug.

Step 7: Transform and Store Data

Add a Transform node to reshape or reformat data as it moves through the workflow, and a State node to save information for later steps.

A common pattern is turning unstructured input into clean JSON, like converting a support message into {"customer_name": "John Doe", "order_id": "12345", "issue": "delivery delay"}, then storing it in a State node for downstream use.

Step 8: Preview and Test

Click Preview in the top-right corner to open a chat window and interact with your agent in real time. The preview pane shows intermediate steps, including which nodes fired and how the agent reasoned, which makes debugging far easier.

Test edge cases, not just the happy path. For a support agent, try "I need to return a defective product," "Where is my order #12345?," and "Can I change my shipping address?" to see how it handles variety.

Step 9: Evaluate Performance

Use the built-in evaluation system to measure accuracy against a dataset of expected inputs and outputs. Build 20 or so test conversations covering common cases, then run them and aim for high accuracy on both responses and routing.

One thing to know: the Evals product is being deprecated alongside Agent Builder, so do not build long-term evaluation infrastructure here. Plan to move this into your post-migration stack.

Step 10: Publish and Deploy

When testing looks good, click Publish and pick a deployment method. You can expose the agent as an endpoint through the Responses API, embed it on your site with a ChatKit embed code, or export the workflow as TypeScript or Python.

That export path matters now for a second reason: exporting to code is also the on-ramp to migrating off Agent Builder, which we cover next.

Step 11: Monitor and Iterate

After deployment, watch conversation logs and track metrics like response accuracy, task completion rate, and average conversation length.

Set a monthly review to update instructions, add knowledge, or tighten guardrails. An agent is not a set-and-forget system, and the edge cases show up in the logs.

What to Use Now That Agent Builder Is Going Away

This is the part no other tutorial covers, and it is the one that determines whether your work survives.

With Agent Builder shutting down on November 30, 2026, OpenAI points to two replacements depending on how you want to build.

Use the Agents SDK when your workflow should continue as code and live inside your own application. Use Workspace Agents in ChatGPT when the use case fits natural-language prompting and you want to share the agent with a team without writing code.

ChatKit's advanced integration stays available for the front end in both cases.

On migration itself, be realistic. You can export an existing Agent Builder workflow as Agents SDK code, but the export does not convert your workflow graph or guarantee that every behavior transfers unchanged.

Expect to rebuild and re-test the logic rather than lift-and-shift it. If you are weighing a code-first build anyway, our guide to building agents with the Claude Agent SDK walks through the same loop in a different ecosystem.

The No-Code Alternative: Build a Custom GPT

If you do not need a workflow engine, a Custom GPT is the simpler path, and it is not affected by the Agent Builder shutdown. It lives inside ChatGPT and requires a paid plan.

Open ChatGPT, click Explore GPTs, then Create. Describe what you want in the builder chat, and refine the instructions it drafts for you.

Upload reference files so the GPT can answer from your own content, and enable tools like web browsing or code interpreter if the task needs them. Test it in the preview, then publish or share it with your team.

A Custom GPT will not branch, loop, or call MCP servers the way Agent Builder does. For a configured assistant trained on your docs, though, it covers most of what people actually want when they search for a "ChatGPT agent builder."

What It Costs to Build and Run an Agent

There is no flat "agent" price. Cost comes from a few moving parts, and the model you pick is the biggest one.

• Model and token usage. Every run consumes API tokens, and a heavier model costs more per call than a smaller one like GPT-5-mini. Tool-heavy workflows that loop through several steps add up faster.
• A paid ChatGPT plan. The Custom GPT path requires a paid ChatGPT subscription to build and share.
• Re-platforming. The hidden cost just surfaced in real time. When a tool is deprecated, the work of rebuilding on its replacement is a real expense, not a rounding error.

To keep spend down, pick the cheapest model that meets your quality bar, watch token usage on multi-step workflows, and avoid betting a critical workflow on a tool you may have to rebuild in six months.

Should You Build Your Own Support Agent, or Buy One?

The Agent Builder shutdown is a useful lesson about a real cost. Building your own support agent means owning the workflow, the guardrails, the integrations, and the evals.

It also means rebuilding all of it when the underlying tool changes course, which is exactly what just happened to teams who standardized on Agent Builder.

For a B2B support team, that maintenance churn is the expensive part. The week you spend wiring nodes is small next to the quarter you spend re-platforming after a deprecation notice.

Support leaders running real ticket volume cannot afford to treat their support agent as a side project that breaks on someone else's roadmap.

You only pay when the AI delivers an outcome, so support stops being a tool you babysit and starts being a result you can measure.

Conclusion

You can still create an AI agent with ChatGPT Agent Builder today, and the node-based flow is genuinely quick to assemble. But with a November 30, 2026 shutdown on the calendar, build with the exit in mind.

Plan for the Agents SDK or Workspace Agents in ChatGPT, or skip the build entirely if what you actually need is a reliable support agent.

If you would rather own the outcome than the maintenance, request access to Helply and let support run on a platform that pays for itself.