---
title: How do you govern agents built with OpenAI AgentKit?
url: https://opencontroller.lyzr.com/blog/govern-openai-agentkit-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you govern agents built with OpenAI AgentKit?

**Use AgentKit's own controls for what they do well: Connector Registry for which data sources and MCP connectors are allowed, Agent Builder versioning for change history, Guardrails for in-agent safety checks and Evals for quality. Then add what AgentKit does not provide: an inventory across platforms, a named owner per agent, cost per agent, independent evidence and a way to stop an agent fast.**

OpenAI announced AgentKit in October 2025. It bundles several building blocks. Some were launched in beta, so check each one's current status in OpenAI's documentation before you write policy around it.

## What does each AgentKit component give governance teams?

| Component | What it does | Status at launch | Governance value |
|---|---|---|---|
| Agent Builder | Visual canvas for multi-agent workflows, with preview runs, inline eval setup and full versioning | Beta | Change history per workflow version |
| Connector Registry | One admin panel for data sources across ChatGPT and the API, including Dropbox, Google Drive, SharePoint, Microsoft Teams and third-party MCPs | Beta rollout to some API, ChatGPT Enterprise and Edu customers | Central allow-list of connectors |
| Global Admin Console | Where Global Owners manage domains, SSO and multiple API organisations | Prerequisite for Connector Registry | Ties admin rights to SSO |
| Guardrails | Open-source safety layer for Python and JavaScript; can mask or flag PII and detect jailbreaks | Open source | In-agent input and output checks |
| Evals | Datasets, trace grading, automated prompt optimisation, third-party model support | Generally available | Repeatable quality evidence |
| ChatKit | Embeddable chat interface for agents | Generally available | Front end only; little governance value |

## What should you configure first?

1. **Turn on the Global Admin Console** and bind it to your identity provider. Connector Registry depends on it.
2. **Build a connector allow-list** in Connector Registry. Start with the connectors your approved use cases need and block the rest. Treat each third-party MCP as a new supplier; our guide on [how to govern MCP servers](/blog/how-to-govern-mcp-servers) has a review checklist.
3. **Require Guardrails** on any agent that handles personal data or faces customers. Enable PII masking and jailbreak detection, and log every trigger.
4. **Keep version discipline** in Agent Builder. A version that reaches production should have a named approver who is not the builder.
5. **Attach an eval dataset** to each production agent and re-run it on every version. Trace grading is useful for multi-step workflows.

## What does AgentKit not cover?

AgentKit governs agents built on OpenAI. Most enterprises also run agents on Copilot Studio, Agentforce, Bedrock, Vertex AI or open-source frameworks. The gaps:

- **Inventory across platforms.** Connector Registry covers OpenAI surfaces, not agents elsewhere. You still need to [find shadow AI agents](/blog/how-to-find-shadow-ai-agents) built with raw API keys.
- **Ownership and lifecycle.** There is no built-in rule that every agent has an accountable owner, or that its access is reviewed.
- **Cost per agent.** Usage is reported by organisation and project. Attributing it to a specific agent version needs a gateway or tagging scheme.
- **Independent evidence.** Auditors prefer evaluation and policy evidence that does not come only from the vendor that runs the agent.
- **Runtime stop.** You need a tested way to block or quarantine one agent without taking down a whole project.

Code-first agents built with the OpenAI Agents SDK have the same gaps; see [governing OpenAI Agents SDK agents](/blog/govern-openai-agents-sdk-agents). For ChatGPT-side GPTs and agents, see [governing ChatGPT Enterprise GPTs and agents](/blog/govern-chatgpt-enterprise-gpts-and-agents).

## Which risks should your AgentKit review map to?

The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) is a practical checklist. For AgentKit agents, focus on goal hijack through prompt injection in connector content, tool misuse through over-broad connectors, and identity abuse where an agent acts with more access than the user. Record which control covers each risk: Guardrails, Connector Registry, an external gateway or a manual review.

## How Opencontroller does this

Opencontroller connects to OpenAI Enterprise and Claude Enterprise workspaces as well as Salesforce, SAP, Databricks and HubSpot, and auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. Everything lands in one governed registry of agents, versions, tools and MCP servers. Lyzr Gateway, the LLM Gateway, attributes cost per agent and per version. Multi-turn and red-team evals and LLM judges on production traffic add evidence that sits outside the agent platform. The pattern is described in [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) and [governing agents across platforms](/blog/how-to-govern-ai-agents-across-platforms). [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is Agent Builder generally available?**
It launched in beta in October 2025. Check OpenAI's current documentation for its status.

**Does Connector Registry cover MCP servers?**
Yes. It manages prebuilt connectors and third-party MCPs across ChatGPT and the API.

**Are Guardrails enough to govern AgentKit agents?**
No. They check inputs and outputs inside an agent. Inventory, ownership, cost and a stop path need other controls.

## Key facts

OpenAI AgentKit, announced in October 2025, includes Agent Builder (visual multi-agent workflow canvas with versioning, preview runs and inline evals; launched in beta), Connector Registry (one admin panel for data sources and MCPs across ChatGPT and the API; beta rollout to some API, ChatGPT Enterprise and Edu customers; requires the Global Admin Console), Guardrails (open-source safety layer for Python and JavaScript that can mask or flag PII and detect jailbreaks), ChatKit and new Evals features (datasets, trace grading, automated prompt optimisation, third-party model support). AgentKit governs agents built on OpenAI; enterprises still need a cross-platform inventory, per-agent identity, cost attribution, independent evaluation evidence and a kill path. Opencontroller by Lyzr connects to OpenAI Enterprise workspaces and other platforms; pricing starts from $250K per deployment per year.

## Related

- [How do you govern agents built with the OpenAI Agents SDK?](https://opencontroller.lyzr.com/blog/govern-openai-agents-sdk-agents.md)
- [How do you govern GPTs and workspace agents in ChatGPT Enterprise?](https://opencontroller.lyzr.com/blog/govern-chatgpt-enterprise-gpts-and-agents.md)
- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
