OpenAI Shows Codex Agents at Work

OpenAI released a new research article on June 25, 2026, examining how agents are transforming work, with Codex as the core example. The piece frames agentic AI as a shift from short chat interactions toward delegated, long-horizon tasks that can run independently with tools and environments. The update is notable for developer-tool watchers because it uses Codex adoption data to show how coding agents are moving beyond software engineering into operations, finance, legal, recruiting, and other knowledge-work functions.

Key Features or Updates

OpenAI reports that Codex usage has shifted toward longer tasks, with many users asking agents to complete work estimated to take more than 30 minutes, one hour, or even eight hours of human effort. The article also highlights parallel agent usage and Codex becoming the primary AI tool across multiple internal OpenAI departments.

Impact on Developers

The findings reinforce that coding agents are becoming workflow platforms, not just code-completion assistants. Developers may increasingly coordinate multiple agents, delegate validation and automation work, and support non-developer teammates who use agentic tools for technical execution.

How to use it

Teams can use the research as a benchmark for where agent workflows are headed: longer tasks, more parallelism, and broader access outside engineering. The practical next step is to identify repeatable internal tasks that can be delegated to Codex-style agents with clear instructions, tooling, and review gates.