Introducing AgentKit, new Evals, and RFT for agents

Source: OpenAI Date: 2025-10-06 URL: https://openai.com/index/introducing-agentkit

Summary

OpenAI’s October 2025 launch of AgentKit — a developer toolkit for building agents on the OpenAI platform — alongside new agent evaluation frameworks (Evals) and Reinforcement Fine-Tuning (RFT) for agent use cases. AgentKit provided higher-level abstractions over the Responses API for common agentic patterns: tool use, memory, multi-step planning, and handoffs. The new Evals tooling enabled developers to measure agent reliability systematically. RFT for agents allowed fine-tuning on task completion trajectories, not just response quality — a technically distinct training approach for agentic behavior.

Implications

RFT as the technical differentiator. Reinforcement Fine-Tuning for agents is methodologically significant: it trains models to optimize for completing multi-step tasks rather than producing good individual responses. This mirrors how o1 was trained on reasoning chains rather than direct answers, but applied to agentic task completion. The October 2025 timing placed it just ahead of the ChatGPT Agent launch (July 2025) entering general availability.

Thread: Agentic API platform. Sits in the developer tooling build-out sequence: Responses API, MCP integration, Codex CLI, AgentKit, and Workspace Agents. Each layer added more structure for developers building autonomous systems on top of OpenAI’s models.

Watch: Whether AgentKit’s abstractions were adopted as the standard pattern for OpenAI-native agent development, or whether developers preferred thinner SDK wrappers (like the OpenAI Agents SDK) over the higher-level toolkit.