Introducing AgentKit, new Evals, and RFT for agents

OpenAI is shipping three things at once today, and the common thread is speed — specifically, the speed of getting an AI agent out of a notebook and into something real users can rely on. The bundle includes AgentKit, an expanded version of its Evals framework, and reinforcement fine-tuning (RFT) tailored for agents rather than single-turn models.

AgentKit is the headline act here. Building an agent that calls tools, holds state across multi-step tasks, and recovers gracefully when something goes wrong has always been the unglamorous, time-consuming part of the job. Most teams end up writing the same scaffolding — orchestration, tool-calling logic, error handling — over and over. AgentKit is OpenAI's attempt to hand developers that scaffolding pre-built, so the actual engineering effort goes into the agent's behavior and not the plumbing underneath it.

The Evals expansion matters more than it might sound. Evaluating agents is fundamentally harder than evaluating a chatbot that just answers questions, because agents take actions across multiple steps and the failure modes compound. A wrong tool call three steps in can wreck the whole task even if every individual step looked reasonable in isolation. Broader evals give developers a way to actually measure that kind of degradation before it ships, rather than discovering it from angry support tickets after launch.

Then there's RFT for agents, which is arguably the most consequential piece long-term. Reinforcement fine-tuning lets a model be optimized against a reward signal tied to outcomes, and applying that specifically to agent workflows means the training loop can now reward an agent for successfully completing a multi-step task, not just producing a plausible-looking single response. That is a meaningfully different training target, and it points toward agents that improve at the thing they're actually deployed to do, rather than just sounding fluent while doing it.

Taken together, this release reads less like a flashy new model and more like infrastructure — the kind of unglamorous tooling that decides whether agentic AI becomes a durable product category or stays stuck in demo purgatory.