Image Courtesy: OpenAI

October 7, 2025 — OpenAI has announced AgentKit, a comprehensive suite of tools that lets developers and enterprises build, deploy, and optimize intelligent agents faster and more reliably.

Until now, building AI agents required a mix of orchestration tools, custom connectors, manual evaluation pipelines, and frontend engineering. AgentKit simplifies that entire process by offering an integrated environment with everything developers need in one place.

What’s in AgentKit?

AgentKit brings together three major components to streamline the agent development lifecycle:

1. Agent Builder — Visual Agent Workflow Designer

A drag-and-drop visual canvas for building and versioning multi-agent workflows.

Developers can connect tools, set up logic, and configure guardrails — all visually.

Agent Builder supports:

• Preview runs

• Inline eval configuration

• Full version control for fast iteration

2. Connector Registry — Unified Data Access and Governance

A central hub for managing data and tool integrations across ChatGPT and API workspaces.

Admins can configure built-in connectors like:

• Dropbox, Google Drive, SharePoint, Microsoft Teams

• Third-party connectors via the Model Context Protocol (MCP)

This registry gives enterprises complete visibility and control over how their data connects to agents — all within a Global Admin Console.

3. ChatKit — Embedded Agentic Chat Experiences

ChatKit lets developers easily add chat-based AI experiences directly into apps or websites — with native support for streaming responses, threads, and branded UIs.

Companies like Canva, HubSpot, and LegalOn are already using ChatKit to embed support and productivity agents.

ChatKit is already powering use cases like:

• Developer support bots

• Internal knowledge assistants

• Customer service agents

• Interactive onboarding flows

Measure and Optimize with Enhanced Evals

OpenAI has also expanded its Evals framework with powerful new features to test and optimize agent performance:

• Datasets — Quickly create evaluation sets and expand them with automated or human grading

• Trace Grading — Assess full workflows to locate weak points

• Automated Prompt Optimization — Suggests improved prompts based on grader results

• Third-party Model Support — Evaluate non-OpenAI models inside the same platform

Early adopters like Carlyle report cutting agent development time by 50% and increasing accuracy by 30%.

Reinforcement Fine-Tuning (RFT)

RFT allows developers to fine-tune reasoning models with custom tool calls and custom grading criteria.

• Available now for OpenAI o4-mini

• Private beta for GPT-5 customers

This lets developers teach agents to call the right tools at the right time and optimize for domain-specific performance metrics.

Pricing and Availability

• ChatKit and Evals — Generally available today for all developers

• Agent Builder — In beta

• Connector Registry — Rolling out in beta for API, ChatGPT Enterprise, and Edu users with Global Admin Console access

• All features are included under standard API model pricing

OpenAI also announced plans to release a Workflows API and agent deployment options directly within ChatGPT later this year.