OpenAI · Technologies
OpenAI AgentKit
Build the agent on a canvas, then keep track of what it connects to.
AgentKit is OpenAI’s toolset for building and running agents. Agent Builder gives you a visual canvas for designing and versioning multi-agent workflows, and the Connector Registry pulls every data source into one administration panel across ChatGPT and the API. Both were in beta at the time of writing.
AGENT WORKFLOWS · EXAMPLE ACCOUNT
WORKFLOW
VERSION
CONNECTORS
STATUS
Lead-Qualify
v4
2
LIVE
Doc-Summarise
v7
1
LIVE
Support-Router
v3
3
LIVE
Quote-Assist
v1
4
IN TESTING
Data-Export
v2
5
BLOCKED
Sample account
live · in testing · blocked
In plain terms
Most agent projects die in version two.
Building one is easy. Changing it safely six weeks later is not.
Without it
- Agent logic lives in a prompt somebody edited
- No way to see what changed or roll it back
- Each agent connects to sources its own way
- Nobody has a list of what is plugged into what
With it
- Workflows are designed and versioned visibly
- You can compare versions and go back one
- Connections come from one managed registry
- An administrator can see every source in use
What CG TECH can do with OpenAI AgentKit
Worth using when you expect to have more than one agent.
We map what the agent does as steps before anything is built, including where it hands to another agent and where it stops. The canvas is useful precisely because that logic stops living in one long prompt.
Agents get edited. We set how versions are promoted, who approves a change and how you roll back, which is what turns a prototype into something a business can depend on.
This is the part IT will care about. One panel showing which data sources agents can reach across ChatGPT and the API, so access is granted and removed centrally rather than per project.
If you need one contained agent over Microsoft 365 content, Copilot Studio is usually faster. We will say so rather than build on the platform we happen to be discussing.
How an engagement runs
From one agent to a set somebody can actually administer.
01
Map the steps
We write down what the agent does before we build it.
02
Build and version
It goes on the canvas, with versions and a way back.
03
Register the sources
Connections go through the registry so access stays visible.
04
Set the change rule
We agree who can edit an agent and who approves it.
Questions we hear a lot
Common questions about OpenAI AgentKit
What is OpenAI AgentKit?
AgentKit is a set of tools for building, deploying and improving agents. The two parts most businesses care about are Agent Builder, a visual canvas for creating and versioning multi-agent workflows, and the Connector Registry, which consolidates data sources into one administration panel.
How is it different from the Agents SDK?
The SDK is code. AgentKit adds a visual way to design and version the workflow, plus central administration of the connections. If your agents are built and maintained by developers, the SDK may be all you need. If other people have to see and change them, the canvas earns its place.
What is the Connector Registry for?
It gives an administrator one place to see and control which data sources agents can reach, across both ChatGPT and the API. Without something like it, each project wires up its own access and nobody has the full picture.
Is it generally available?
Agent Builder has been available in beta, and the Connector Registry has been rolling out in beta to some API, ChatGPT Enterprise and Edu customers. Check the current status before planning a rollout around it, as beta features move.
Should we use this or Microsoft Copilot Studio?
It depends where your content and systems live. Copilot Studio is the shorter path for Microsoft 365 and Dataverse. AgentKit suits work spanning OpenAI models and sources outside Microsoft. We assess that before recommending either.
Who should own these agents internally?
Someone named, with time for it. The common failure is an agent built by an enthusiast who changes roles. We set ownership and a review date as part of delivery.
Can you work with agents we have already built?
Yes, and that is often the more useful engagement. We review what is running, what it connects to and what happens when it fails, then put version control and ownership around it.
How do we stop agents reaching data they should not?
Central connector administration is half of it. The other half is fixing permissions and oversharing in the source systems, because an agent generally inherits the access it is given. Both need doing.
Ready when you are
Building agents faster than you can keep track of them?
That is the point where this matters. A discovery session covers what you have running now, what it connects to, and how to get it under proper change control.
What to expect
- A consultant replies within 4 business hours
- Session booked to understand your requirements
- We will provide you with a fixed price quote