OpenAI Agent Builder alternative: FlowRunner vs OpenAI Agent Builder

TL;DR

• OpenAI shuts down Agent Builder on November 30, 2026. The Evals dashboard and API and the reusable prompts API (v1/prompts) close the same day, and existing evals have been read-only since October 31. ChatKit stays.
• After November 30 you can’t open a workflow to copy its code, so export every one now. Your vector stores stay in your OpenAI account, and the export carries only their IDs.
• OpenAI offers two paths: run the Agents SDK export in your own application, or rebuild the workflow as a ChatGPT workspace agent. FlowRunner is a third: the same agents, instructions, and models on your own OpenAI key, as a flow that runs without a server of yours.
• In our tests the export turned human approval into a function that always returns true. On FlowRunner, the approval reaches a person on email, Slack, WhatsApp, phone, or a Form and waits for the answer, up to 30 days on Free, Starter, and Growth and up to a year on Professional and above.

What closes, and when

OpenAI’s deprecations page has three entries with the same end date.

• Agent Builder shuts down on November 30, 2026. ChatKit remains available.
• Evals: existing evals became read-only on October 31, 2026. The dashboard and API shut down on November 30. Save the results you want to keep.
• Reusable prompts (v1/prompts) shut down on November 30. OpenAI’s advice is to move that prompt content into your application code.

What that means for what you built:

• The workflows. Once Agent Builder is gone you can’t open them. The only copy you keep is the code you export: open each workflow, click Code, choose the Agents SDK tab, and pick TypeScript or Python.
• Your files and vector stores. File search points at vector stores in your OpenAI account. They stay there, and the export carries only their IDs.
• Documents uploaded in the chat. They aren’t part of the exported input.

Who this comparison is for

You built agent workflows in Agent Builder and need them running somewhere by November 30. OpenAI’s migration guide gives you two options:

• The Agents SDK. Run the exported code in your own application. You host it, keep it running, and build whatever the canvas used to do for you.
• ChatGPT workspace agents. Create an agent from the export inside ChatGPT. That needs a ChatGPT Business, Enterprise, or Edu workspace.

OpenAI’s guide is also clear that the export doesn’t convert your workflow graph or guarantee that every behavior transfers unchanged.

This page compares a third option. Start with what Agent Builder and FlowRunner share: a canvas where agents with instructions, models, and tools are wired together with branches, guardrails, and an approval step. FlowRunner does that too. It also runs the flow on its own, started by an event or a call, reaches people outside the chat when a step needs them, and keeps a record of every run. If you’d rather keep the SDK code, FlowRunner can sit beside it and handle the approvals.

What FlowRunner is good at

FlowRunner is built around agents that know when to stop and ask a person. Human-in-the-loop is an execution pattern where AI agents pause autonomously, assemble the relevant context and the decision choices available, route to a human via their preferred channel, and resume the moment the human responds.

• Agents as steps in a flow. The AI Agent block keeps your instructions, your model, and your structured output, and later steps read each output field by name. Agents run on your own provider key with no markup (BYOK).
• Approvals that wait. A human step sends the request on email, Slack, WhatsApp, phone, or a Form, waits up to 30 days on Free, Starter, and Growth and up to a year on Professional and above, and continues with the answer. An agent can call that step as one of its tools and suspend until it returns.
• Many ways to start. A flow can start from the Call Flow API, a webhook, a schedule, an incoming email, an event in one of 2,100+ integrations, or another flow.
• Tools for agents. 56,000+ callable actions (as of September 2026), each with the metadata an agent needs to use it, plus your other flows, MCP servers you register, Knowledge Bases, and web search services.
• A record of every run. Each run is stored block by block, with each block’s input and output, and anyone in the workspace can open it. A failing block can route to a Handle Error path instead of ending the run.
• Nothing to host. Flows run in FlowRunner’s cloud, with unlimited users and unlimited workflows on every plan.

What OpenAI Agent Builder is good at

Agent Builder is OpenAI’s visual canvas for agent workflows on the OpenAI platform. A workflow starts from a chat message and runs through nodes: Agent, File search, Guardrails, MCP, If/else, While, Human approval, Transform, and Set state. You test it in Preview, check it with Evals, and put it in front of users with ChatKit or the exported SDK code. It has no fee of its own: you pay for model tokens and built-in tools at standard API prices.

Where FlowRunner is stronger

• Approvals that really wait. Agent Builder’s Human approval node asks the person in the chat, and OpenAI documents no limit on how long it waits. In our export of six OpenAI templates, that node came out as a function that always returns true, so on the SDK path nobody is asked until you build the approval again. The Agents SDK has its own approval pause, but its docs leave storing the paused run, notifying the reviewer, and checking who approved to you. On FlowRunner, the approval is a step that reaches the person and waits up to 30 days on Free, Starter, and Growth or up to a year on Professional and above, and the run counts as one execution however long it waits.
• Runs without a chat and without a server. An Agent Builder workflow starts from chat text and goes live through ChatKit or code you host. A FlowRunner flow runs in FlowRunner’s cloud, starts from whatever event should start it, and returns its answer to whoever called it. If it needs a person, it goes and finds them.
• The approver doesn’t need an account. The person answers from an email, a Slack message, WhatsApp, a phone call, or a Form’s public link. Nobody signs in to FlowRunner to approve.
• More reach for your agents. Agent Builder’s MCP node connects OpenAI connectors and third-party servers. FlowRunner agents get 2,100+ integrations as tools, plus any MCP server you register, and the other flows in your workspace.
• Your choice of model. The converted agents keep their OpenAI models on your key. Any agent can switch to another provider later, one block at a time.
• A record that outlasts the tool. Evals is closing with Agent Builder. In FlowRunner, every run stays in run history with each block’s input and output, and the Professional plan adds audit trails and RBAC at $299 a month.

Where OpenAI Agent Builder is stronger

• OpenAI’s hosted tools on the node. An Agent Builder agent can use Code Interpreter, which runs code the model writes during the run, and file search over vector stores in your OpenAI account. FlowRunner’s Custom Cloud Code runs JavaScript written ahead of time, and file search becomes a Knowledge Base you load your documents into. Both of these Agent Builder tools also sit poorly in the export: file search comes out as vector store IDs only, and in our conversion of OpenAI’s Internal Knowledge Assistant template, the fact-finding agent was given web search in place of Code Interpreter.
• A chat interface out of the box. Agent Builder workflows run inside a chat, and ChatKit embeds that chat in your site. ChatKit stays after November 30, but its tab in Agent Builder gives you an ID that points at the Agent Builder workflow, which won’t help once the workflow is gone. A FlowRunner flow returns its answer to whatever calls it, so a chat front end calls the flow.

Starting free

Agent Builder had no fee of its own. OpenAI billed the model tokens your agents used and the built-in tools they called.

That part doesn’t change on FlowRunner. Your agents call OpenAI with your key, at your rates, on your OpenAI invoice. FlowRunner adds no AI markup and counts one thing: the execution, one complete run of a flow. A run with eight agents counts once, and so does a run that waits two days for an approval.

The Free plan is $0, permanent, and needs no card: 100 executions a month, one at a time, 24 hours of log history, unlimited users and workflows, every integration, AI agents, and human-in-the-loop. Every new account also gets 14 days of Professional on its first workspace. Above Free, Starter is $5 a month for 300 executions or $15 for 3,000, and Growth starts at $45 for 12,000.

Feature comparison

Dimension	FlowRunner	OpenAI Agent Builder
Status	Active	Shuts down November 30, 2026
Where it runs	FlowRunner cloud; nothing to host	OpenAI platform; deployed through ChatKit or exported SDK code
How a run starts	Call Flow API, webhook, schedule, incoming email, integration event, another flow	A chat message (input_as_text)
Agents	AI Agent block: instructions, model, structured output	Agent node: instructions, model, structured output
Models	Your own key for OpenAI or another provider, no markup	OpenAI models, billed by OpenAI
Human approval	Reaches the person on email, Slack, WhatsApp, phone, or a Form	Human approval node, answered in the chat
Approval in the code export	Not applicable; the flow keeps the real step	A function that always returns true (our test)
Longest approval wait	30 days on Free, Starter, Growth; 1 year on Professional and above	Not documented
Guardrails	Condition with an AI Question	Guardrails node (moderation, jailbreak, PII)
Branching	Condition, Value Router, AI Router	If/else
Loops	Repeat, List Iterator	While
State	Set Variables; Shared Memory across runs	Set state
Document search	Knowledge Bases	File search over OpenAI vector stores
Web search	Web search services attached to the agent as tools	Built-in Web Search (breaks the export)
Code	Custom Cloud Code, JavaScript, also an agent tool	Code Interpreter
MCP	Agents use tools from registered MCP servers	MCP node: OpenAI connectors or third-party servers
Integrations	2,100+ integrations, 56,000+ callable actions	Through MCP and function tools
Run record	Every run stored block by block, with inputs and outputs	Preview and traces; Evals (closing)
Billing	One execution per complete run; AI on your own key	Model tokens and built-in tools at API prices

Decision framework

Pick FlowRunner if:

• You want your Agent Builder workflows running by November 30 without standing up a service to host them.
• A workflow has an approval, and the person should get the request on their own channel and have days to answer.
• The workflow should start from a business event or a schedule, not only a chat.
• You want your OpenAI models on your own key, and the option to move an agent to another provider.
• Teammates outside engineering need to read the flow and open its runs.

Pick the Agents SDK if:

• You want the exported code as your starting point and you’re set up to host and maintain it.
• You’re ready to build the approval loop yourself: store the paused run, notify the reviewer, and resume. Or call FlowRunner for that part.

Pick ChatGPT workspace agents if:

• Your team works in ChatGPT Business, Enterprise, or Edu, and the agent only needs to work inside ChatGPT.

Ready to move?

You don’t rebuild your workflows by hand. The Agent Builder migration guide shows how FlowRunner Prompt to Flow converts one workflow at a time from its exported code, and what’s left for you to connect and test before November 30. For a bigger migration, contact us.