AgentKit provides an integrated suite that lets developers create, manage, and improve AI agents without juggling disconnected tools.

Agent Builder – visual workflow designer

Agent Builder offers a drag‑and‑drop canvas where teams can map out multi‑agent logic, add guardrails, and version each iteration. The interface promotes rapid prototyping and clear communication across product, legal, and engineering groups.

• Canvas supports nodes for tool calls, conditionals, and loop constructs.
• Built‑in preview runs let you test a workflow instantly.
• Full version history enables safe rollback and A/B testing.
• Template library accelerates common patterns such as support bots and research assistants.
• Integrates with large language models via the Responses API.

Connector Registry – centralized data and tool management

The Connector Registry aggregates external services into a single admin panel, allowing administrators to control access and configure connections for all agents in an organization.

• Pre‑built connectors for Dropbox, Google Drive, SharePoint, Microsoft Teams, and more.
• Custom connector framework supports third‑party MCPs.
• Global Admin Console governs domains, SSO, and multi‑org API keys.
• Guardrails can be attached at the connector level to mask PII or block jailbreak attempts.
• Works within Choosing the right AI model guidelines.

ChatKit – embedable chat‑based agent UI

ChatKit reduces the effort required to integrate conversational agents into web or mobile products, handling streaming responses, thread management, and UI theming.

• SDKs for JavaScript and React simplify embedding.
• Customizable themes align the chat appearance with brand guidelines.
• Supports real‑time “thinking” indicators to improve user trust.
• Built‑in analytics capture interaction metrics.
• Compatible with cloud computing architecture for scalable deployments.

Evals Enhancements – measuring and improving agent performance

New eval capabilities let developers create datasets, run trace grading, automate prompt refinement, and benchmark third‑party models, all from a single interface.

• Dataset creator with automatic grader attachment.
• Trace grading visualizes end‑to‑end workflow execution.
• Prompt optimizer generates higher‑quality prompts based on human feedback.
• Third‑party model support expands evaluation beyond OpenAI offerings.
• Integrated with secure development environments best practices.

Reinforcement Fine‑Tuning – custom reasoning for specific tasks

Reinforcement fine‑tuning (RFT) allows teams to adapt reasoning models to call the right tools at the right moment and to enforce custom success criteria.

• Custom tool‑call training improves workflow efficiency.
• Custom graders let you define task‑specific evaluation metrics.
• Available on o4‑mini and in private beta for GPT‑5.
• Beta feedback loop informs future model releases.
• Works in conjunction with Agent Builder guardrails for safe operation.