OpenAI has launched AgentKit, a new suite of tools for building, deploying, and evaluating Artificial Intelligence (AI) agents, which are systems that can autonomously perceive information, make decisions, and take actions to achieve specific goals.
Traditionally, building production-ready agents required managing multiple disconnected tools (eg. custom connectors, manual evaluation pipelines, frontend development, etc.), which AgentKit now consolidates into a single interface.
Key capabilities of AgentKit
1. Agent Builder
AgentKit’s Agent Builder provides a visual canvas for assembling and managing multi-agent workflows. Teams can create logic flows using drag-and-drop nodes, integrate external tools, and set up guardrails directly within the interface. The platform also allows users to test and refine agents’ behaviour in real time before deployment.
Guardrails, an open-source safety layer designed to prevent unintended or unsafe behaviour, can also be enabled in Agent Builder. They can detect sensitive data, flag jailbreak attempts, and ensure compliance with other safeguards. Guardrails may be deployed independently or integrated via the Guardrails library for Python and JavaScript.
Learn more about Agent Builder here.
2. Connector Registry
The Connector Registry enables enterprise administrators to manage how data and APIs are connected across OpenAI products such as ChatGPT and the API platform in a centralised administrator panel. It supports pre-built connectors (eg. SharePoint, Teams) and third-party Model Context Protocols (MCPs).
Learn more about the Connector Registry here.
3. ChatKit
ChatKit provides a toolkit for embedding customisable conversational interfaces to websites or applications. It manages core chat functions such as message streaming, conversation threading, and response handling. This enables developers to add chat interfaces to their products without extensive frontend development.
Read more about ChatKit here.
Complementary Updates
Alongside the launch of AgentKit, OpenAI has also introduced several complementary updates to its broader developer system. While not exclusive to AgentKit, these enhancements can be applied to the overall toolkit to support AI agent building and refinement.
1. Expanded Evaluation (Eval) Capabilities
OpenAI has introduced four new capabilities to its Evals toolkit, which aims to support developers in testing and refining their agents. These enhancements are:
- Datasets: Build and expand evaluation sets over time.
- Trace grading: Analyse end-to-end agent workflows for performance issues.
- Automated prompt optimisation: Generate improved prompts based on human and machine feedback.
- Third-party model evaluation: Assess models from other providers within the same framework.
Read more about Evals here.
2. Reinforcement Fine-Tuning (RFT)
OpenAI’s RFT capability, now generally available for o4-mini and in beta for GPT-5, allows developers to tailor model reasoning for specific use cases. The new update introduces:
- Custom tool calls: Training agents to select the appropriate tools for a given task.
- Custom graders: Enables teams to define evaluation criteria aligned with organisational priorities.
Learn more about RFT here.
AgentKit Availability for ChatGPT Enterprise Users at Mediacorp
AgentKit is currently under review by Mediacorp’s AI Council. The AI Team will provide updates when it is available to staff.
References
