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·Apr 15·5 min read
The next evolution of the Agents SDK
TL;DR
Building long-running autonomous agents (software that operates independently over time) required developers to manually handle sandboxing (isolated execution environments) and tool integration, creating security and reliability gaps. OpenAI released an updated Agents SDK with native sandbox execution and a model-native harness (built-in framework for connecting AI models to tools), eliminating manual setup.
✦ Why It Matters
You can now build and deploy secure, long-running agents without writing custom sandboxing or tool-integration code.
Key Takeaways
- Use native sandbox execution to isolate agent code and prevent unauthorized system access automatically.
- Leverage the model-native harness to connect AI models to external tools without custom integration code.
- Deploy agents that persist across multiple file operations and tool calls without manual infrastructure setup.
- Reduce security risk and development time by adopting built-in sandboxing instead of custom isolation layers.
Full Summary
Autonomous agents—AI systems that operate independently to complete multi-step tasks—have grown in capability but remained difficult to deploy safely at scale. Previous approaches required developers to manually implement sandboxing (isolated runtime environments preventing unauthorized system access) and build custom bridges between language models and external tools, increasing complexity and security risk.
OpenAI's updated Agents SDK introduces native sandbox execution, which automatically isolates agent code in secure containers, and a model-native harness, a framework that natively integrates AI models with external tools and APIs without custom glue code. This enables developers to build agents that can read and write files, call multiple tools sequentially, and run for extended periods while maintaining security guarantees.
The improvements reduce boilerplate code and lower the barrier to deploying production-grade agents, making long-running autonomous workflows more accessible to engineering teams.