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·Mar 11·5 min read
From model to agent: Equipping the Responses API with a computer environment
TL;DR
Language models like GPT can generate text but lack ability to execute code, access files, or maintain state—limiting real-world automation. OpenAI built an agent runtime combining the Responses API (structured output format), shell tool execution, and hosted containers to enable secure, stateful agent operations.
✦ Why It Matters
You can now deploy autonomous agents that execute code and manage files securely without building custom infrastructure.
Key Takeaways
- Use Responses API for structured, parseable model outputs in agent systems.
- Isolate agent execution in containers to enforce security and resource limits.
- Combine language models with file system and shell access for autonomous workflows.
- Design agents as decision-makers that generate commands for external execution.
Full Summary
Language models excel at text generation but cannot natively execute code, manipulate files, or maintain persistent state across sessions—gaps that prevent them from functioning as autonomous agents. OpenAI constructed an agent runtime architecture leveraging three core components: the Responses API (a structured output mechanism ensuring predictable, parseable model responses), a shell tool for command execution, and containerized hosting environments for isolation and scalability.
The approach treats the model as a decision-maker that generates structured commands, which the runtime interprets and executes in a sandboxed container with file system and shell access. This design enables agents to perform multi-step tasks—reading inputs, executing logic, modifying state, and returning results—while maintaining security through container isolation and resource limits.
The system demonstrates that coupling language models with execution environments and persistent storage transforms them from text generators into practical autonomous agents capable of real-world automation tasks.