We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Oct 6·5 min read
Introducing AgentKit, new Evals, and RFT for agents
TL;DR
Developers faced friction moving AI agents from experimental prototypes to production systems, lacking integrated tooling and evaluation methods. OpenAI released AgentKit (a developer framework), expanded evals capabilities (testing tools), and RFT for agents (reinforcement fine-tuning technique).
✦ Why It Matters
Engineers can now build, test, and optimize production agents faster using integrated OpenAI tools instead of custom infrastructure.
Key Takeaways
- Use AgentKit framework to structure agent development and reduce custom scaffolding work.
- Implement expanded evals to systematically test agent behavior before production deployment.
- Apply RFT to improve agent performance using feedback from real-world interactions.
- Combine all three tools to accelerate iteration cycles from prototype to production systems.
Full Summary
Building AI agents that work reliably in production requires more than a working prototype—developers need frameworks to manage complexity, ways to measure agent behavior, and techniques to improve performance. OpenAI released three complementary tools addressing this gap.
AgentKit provides a structured framework for building agents with integrated utilities. Expanded evals capabilities enable developers to systematically test agent behavior across scenarios, similar to unit testing in traditional software.
Reinforcement fine-tuning (RFT) for agents applies reinforcement learning—training systems to maximize rewards—to improve agent decision-making based on real-world feedback. Together, these tools streamline the prototype-to-production pipeline by providing scaffolding, measurement, and optimization mechanisms that were previously manual or absent.