OpenAI has added a workflow to Codex that lets the agent watch a user perform a task once and then repeat it in the future. The feature is meant to reduce the friction of turning recurring developer work into repeatable automation.
That could make coding agents more useful outside of one-off code generation. If an agent can learn local project routines, it can help with repetitive maintenance, testing, data cleanup, or release preparation without requiring users to write full scripts first.
The hard part will be control and verification. A remembered workflow is only useful if developers can inspect what the agent learned, constrain when it runs, and catch drift when the project changes.