Agent Builder
A visual canvas for multi-step workflows. Nodes include agents, file search, guardrails, MCP tools, if/else and while logic, and human approval. OpenAI is deprecating it, and it is scheduled to shut down on November 30, 2026.
Compare
OpenAI AgentKit
a set of tools from OpenAI for building, deploying, and evaluating agents, including the visual Agent Builder, ChatKit, and the Agents SDK
Who they are and what they are good at
OpenAI launched AgentKit in October 2025 for developers and product teams building agents on the OpenAI platform. Its parts are moving in different directions.
ChatKit
An embeddable chat interface for putting a published workflow inside your own product. OpenAI says it remains available.
Agents SDK
A code framework whose runner handles the agent loop and handoffs. An Agent Builder workflow can be exported as SDK code to run on your own infrastructure.
Evals
Trace grading and test datasets. The Evals platform is also being retired: existing evals become read-only on October 31, 2026, and the dashboard and API shut down on November 30.
AgentKit is good at what it was built for: engineering and product teams putting a chat-based agent in front of customers or employees, with OpenAI models, guardrails, and testing close at hand.
Where AgentKit stops for messy back-office work
A recurring back-office procedure needs more than a workflow graph.
The canvas is going away
OpenAI points Agent Builder users to the Agents SDK for code, or to workspace agents in ChatGPT for no-code team workflows. A procedure your team depends on should live somewhere you control.
No screen tool on the canvas
Agent Builder's nodes reach software through file search and MCP. There is no computer use node, so a desktop ERP or a portal with no API needs another route.
The SDK is a kit
Export gives you code. Hosting, logins, scheduling, review screens, and the evidence store become your engineering project.
Exact math
Tying an accrual to the general ledger to the cent should be a rule, not a model's judgment.
LaunchAI is built by the person who does the job, operates any app through the screen, and keeps a screenshot of every action. See Product →.
Side by side
| AgentKit | LaunchAI | |
|---|---|---|
| What it is | Developer tools for chat-based agents | A desktop product for recurring procedures |
| Visual builder | Agent Builder, until November 30, 2026 | The Studio: talk and show, each step checked live |
| Who builds | Developers and product teams | The process owner |
| How it reaches software | File search, MCP, and your code | The screen, through the Vision Layer; no APIs needed |
| Where it runs | OpenAI's platform, or your servers with the SDK | The user's machine or an enrolled agent PC |
| Desktop apps with no API | Not from the canvas | Any app on the machine |
| Exact math | Model judgment or your code | Deterministic rules, no model call |
| Human review | A human approval node | A review step anywhere in the workflow |
| Evidence | Traces | Before-and-after screenshots of every action |
| Model choice | OpenAI | OpenAI, Anthropic, Azure, Bedrock, or self-hosted, per step |
| Cost unit | API usage | License, plus your model account |
When to choose AgentKit
- A software team wants a support agent inside its own app, built in code with the Agents SDK and shown through ChatKit.
- A team moving off Agent Builder has engineers who want to own the exported code.
- For no-code team agents inside ChatGPT, OpenAI now points to workspace agents. See OpenAI agents →.
When to choose LaunchAI
- Month-end accruals: pull open purchase orders from a desktop ERP, match receipts, tie the total to the ledger to the cent, and stop for the controller before the journal entry posts.
- The procedure runs on a schedule, on screens with no API, under the accountant's own login.
- The process owner should be able to change the agent without a developer.
Better together
LaunchAI's model steps, like reading a vendor invoice or finding a moved field, can run on your OpenAI or Azure OpenAI account. Deterministic steps make no model call. Keep the Agents SDK and ChatKit for the agents your product team ships to customers.
FAQ
Is Agent Builder going away?
Yes. OpenAI notified developers on June 3, 2026 that it is deprecated and scheduled to shut down on November 30, 2026. ChatKit remains available.
Where does OpenAI suggest moving workflows?
To the Agents SDK for code, or to workspace agents in ChatGPT for no-code use. Agent Builder can export a workflow as Agents SDK code.
Can an Agent Builder workflow click through a desktop app?
Not with its built-in nodes. It reaches software through file search and MCP tools. Screen work needs your own computer use setup or a product built for it.
We built an AP workflow in Agent Builder. What carries over to LaunchAI?
The knowledge: which systems, which fields, which checks, and where a person approves. Rebuilt as small steps in the Studio, each step is checked live as you build.
Can LaunchAI use our OpenAI account?
Yes. You can pin a model per step, and deterministic steps make no model call.
No prompts. No babysitting. No Integrations.