On This Page8 sections
Key Takeaways
- OpenAI AgentKit is a collection of tools for building, connecting, deploying, and evaluating AI agents. OpenAI introduced it in October 2025. The model supplies reasoning, while AgentKit helps developers define what the agent can do and how people use it.
- Agent Builder is scheduled to shut down on November 30, 2026. As of November 5, existing users can still export their workflows. ChatKit remains available for embedding chat in applications. OpenAI’s deprecation timeline
- For a new customer-facing app, the OpenAI Agents SDK can run the workflow in the application’s backend, with ChatKit providing the interface. For a team workflow that belongs inside ChatGPT, OpenAI also documents a Workspace Agents migration path. Official migration guide
What is AgentKit?
OpenAI AgentKit brought several jobs under one name. A developer could draw the steps of a workflow, connect it to tools and data, publish a version, embed it in a website, and inspect how it behaved on test cases. The underlying model could change without changing the purpose of the workflow. OpenAI’s developer overview
Consider a career adviser. A model can suggest courses from a prompt. An agent can first read an uploaded résumé, identify which skills a particular role requires, search for courses that teach those skills, and return links. AgentKit gave developers a way to specify those steps and inspect the output of each one. That matters when the final answer looks plausible but recommends a course for a skill the applicant already has.
The name also appears in unrelated developer tools. A tutorial that imports @coinbase/agentkit concerns wallets and blockchain actions. One that imports @inngest/agent-kit uses Inngest’s framework. Checking the import line is a quick way to identify which tutorial applies. Coinbase package reference · Inngest agent reference
What did OpenAI AgentKit include?
| Component | What it did | Why it mattered |
|---|---|---|
| Agent Builder | Connected workflow steps on a visual canvas. Each connection defined the data the next step expected. Publishing created a versioned snapshot. | A team could inspect where a run went wrong and identify which workflow version produced the result. |
| Connector Registry | Gave administrators a central place to manage connections to external tools and data, including Model Context Protocol (MCP) servers. | Teams could govern which integrations were available. An application still had to enforce each user’s permissions when accessing records. |
| ChatKit | Embedded a chat interface with features such as file attachments and interactive widgets. | Developers could change the server that ran the agent while keeping a familiar interface for users. |
| Evaluation tools | Used datasets and trace grading to assess answers and the steps that produced them. | A test could catch an agent that reached the expected answer after calling the wrong tool. |
The components have different lifecycles. Alongside Agent Builder’s shutdown, existing OpenAI Evals became read-only on October 31, 2026. The hosted Evals dashboard and API are scheduled to close on November 30. OpenAI documents how to recreate test cases and grading criteria in Promptfoo so they can run locally or in continuous integration. Agent Builder documentation · Evals transition dates · Promptfoo migration guide
How AgentKit works in a real example
In an OpenAI Cookbook walkthrough, a teacher uploads a résumé and asks how to become a school superintendent. The example uses three agents in sequence. The first extracts skills and supporting evidence. The second compares those skills with the target role. The third searches the web for courses that address the gaps. ChatKit presents the recommendations to the applicant.
The most consequential design choice is what reaches the search step. That step needs a skill and a learning goal. It has little use for the applicant’s name, phone number, or complete employment history. A developer adapting the example could restrict the search input to a record like this:
{
"target_role": "school superintendent",
"skill_to_research": "district budget planning"
}That boundary also makes failures easier to diagnose. Suppose a test résumé explicitly says the applicant managed a district budget, yet the app recommends introductory budgeting courses. Inspect the extracted skills first. If budgeting is missing, fix extraction. If the evidence appears there, inspect the gap analysis. Changing the search prompt would leave either upstream error in place.
A second test can place an instruction inside the résumé asking the agent to send the full document to a website. The expected outcome is observable: the search tool receives only the approved fields, and the document stays out of the tool call. Validate those fields in application code and inspect the actual arguments during testing. OpenAI’s guidance on structured outputs and untrusted input
AgentKit vs. the Agents SDK, Agents API, and ChatKit
The practical question for a new project is who should run the workflow? ChatKit answers a separate question about what the user sees. OpenAI’s agent runtime comparison
| Project requirement | Starting point | Practical consequence |
|---|---|---|
| An application needs its own approval rules, tools, or handoffs between agents | Agents SDK | The SDK runs the agent loop inside the application. The team controls deployment, permissions, and storage. |
| A long-running task needs OpenAI to manage execution and preserve progress between turns | Agents API | OpenAI runs the managed Codex harness. The developer configures the task, tools, and execution environment. |
| A feature needs a direct model response, or the application already controls each step | Responses API | The application can call the model directly and decide how much orchestration to add. |
| A website needs an embedded conversation with widgets or attachments | ChatKit | For a new integration, connect ChatKit to an agent backend on the team’s own infrastructure. |
For the career adviser, one résumé summary might justify a direct Responses API call. Adding skill analysis, web search, and a rule about which data may leave the application makes the Agents SDK useful. A scheduled batch job can use that same workflow without any chat interface. ChatKit integration options
How much does AgentKit cost?
Estimate the cost of a completed task. In the career example, one user request can trigger three model steps, a web search, and a retry if a course lookup fails. Track input and output tokens, paid tool calls, and storage separately. Divide the total spend by the number of requests that actually produced usable recommendations.
For a concrete example, OpenAI lists standard web search at $10 per 1,000 calls, with search-content tokens billed at model rates. If 1,000 career plans each trigger two searches, that is $20 in search-call fees before token charges. OpenAI also lists ChatKit file and image upload storage at $0.10 per GB-day after its stated free allowance of 1 GB per account per month. An application that stores uploads itself should budget for its own storage provider. OpenAI API pricing
How to migrate an existing Agent Builder workflow
- Export the workflow while the canvas is available. Open the workflow, select Code, choose Agents SDK, then copy the complete Python or TypeScript export. Record which published version the live app uses so there is a clear baseline for comparison.
- Choose where it will run. For an application, configure the matching SDK in the backend. For an internal workflow, a ChatGPT Business, Enterprise, or Edu workspace with the necessary access and creation permissions can use the export to recreate a Workspace Agent. Review any branch that must execute in an exact order.
- Reconnect tools and check permissions. Test with an ordinary user account as well as an administrator. For a tool that can change records, check what happens if the tool succeeds but the response times out. A retry should recognize the completed action and avoid applying it twice.
- Compare behavior before switching traffic. Replay a successful task, a denied request, a failed lookup, and an input containing a misleading instruction. Compare the chosen tools, their arguments, and their effects alongside the final answer. If the frontend points to an Agent Builder workflow ID, connect it to the new backend and test a full conversation there.
An export gives the team workflow code. Credentials, permissions, deployment settings, and behavior in the new runtime still require verification. OpenAI’s Agent Builder migration steps · Self-hosted ChatKit guide
Conclusion
For a team already using Agent Builder, the most valuable step before November 30, 2026 is to export one live workflow and replay a case with a known outcome, including the tool calls it should make. That comparison gives the team something concrete to verify before moving users.
For a new project, start with one task and write down what a correct result looks like. Use the Responses API if one model call can do the job. Choose the Agents SDK when the application needs to coordinate tools or enforce approvals, and add ChatKit when users need a conversation. That decision can be made from the shape of the task itself.
Continue Reading
More articles connected to the same themes, protocols, and tools.
Referenced Tools
Browse entries that are adjacent to the topics covered in this article.
