Token Robin Hood
OpenAI CodexApr 18, 20268 min

OpenAI Codex expands beyond coding: computer use, automations, memory, and agent workflows

OpenAI's April 16 Codex release turns Codex from a code assistant into a broader software-workflow agent. It can operate a Mac with its own cursor, use an in-app browser, generate images, remember preferences, run automations, and connect more deeply to developer tools.

What happenedCodex gained computer use, plugins, image generation, memory preview, automations, SSH, PR review support, and multi-file workspace improvements.
Why builders careAgent work is moving from code edits into the full software lifecycle.
TRH actionGive Codex clear scope, output contracts, and source limits before letting it operate across apps.

What OpenAI announced

OpenAI says Codex now serves more than 3 million weekly developers and is being updated to work across the full software development lifecycle. The app can use background computer control on macOS, work with an in-app browser, generate and iterate on images with gpt-image-1.5, and use more than 90 additional plugins across tools such as GitLab Issues, CircleCI, CodeRabbit, Render, Microsoft Suite, Atlassian Rovo, Remotion, Neon, and Superpowers.

Why this is more than a product feature drop

The important shift is that Codex is moving from isolated code tasks toward operating real workflows: review comments, terminals, PDFs, spreadsheets, slides, docs, remote devboxes over SSH, visual iteration, and scheduled follow-up work. OpenAI also says thread-based automations can preserve context across days or weeks, while memory can preserve preferences, corrections, and hard-won project facts.

The risk: broader context means broader waste

The same features that make Codex more useful can also make agent runs more expensive and harder to audit. A computer-use agent can click through irrelevant UI. A memory-enabled agent can carry stale assumptions. A plugin-rich agent can gather more context than the task needs. A visual workflow can burn tokens on screenshots and repeated inspection loops.

TRH operating playbook

Before delegating a workflow to Codex, define the target artifact, allowed apps, source boundaries, stop conditions, and verification command. Use automations for recurring discovery and drafting, but keep publishing gates human-approved. For teams, track the ratio between context gathered, actions taken, and artifacts shipped. That is the difference between agent leverage and agent drag.

Sources