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·Jan 23·8 min read
Unrolling the Codex agent loop
TL;DR
Codex, OpenAI's code-generation model, required a way to orchestrate multiple components—models, external tools, and prompts—into a cohesive agent loop. OpenAI built Codex CLI, a command-line interface that chains Codex with the Responses API to enable iterative code generation and tool use.
✦ Why It Matters
Engineers can build autonomous code agents that iteratively refine solutions using model outputs and real-time tool feedback.
Key Takeaways
- Use Codex CLI to coordinate model calls with external tool execution and feedback loops.
- Structure agent loops as: generate → execute tool → observe → refine for iterative problem-solving.
- Leverage Responses API to standardize how models receive tool outputs and format follow-up prompts.
- Separate model reasoning from tool execution to enable scalable, debuggable code-generation workflows.
Full Summary
Codex is OpenAI's large language model trained on code, but deploying it as an autonomous agent—a system that reasons about tasks, takes actions, and learns from outcomes—requires orchestrating multiple moving parts. The Codex agent loop is a feedback cycle: the model generates code or reasoning, calls external tools (APIs, linters, test runners), observes results, and refines its approach.
OpenAI developed Codex CLI, a command-line tool that implements this loop using the Responses API, which standardizes how the model receives tool outputs and formats subsequent prompts. The architecture separates concerns: the model handles reasoning, tools handle execution, and the CLI manages state and sequencing.
This design allows engineers to build reliable code-generation pipelines where Codex can autonomously debug, test, and iterate. The approach demonstrates how to scale agent behavior beyond single model calls, enabling more complex software engineering tasks.