OpenAI's Codex Scratchpad: Parallel Agents Are Becoming Table Stakes

This week, OpenAI tipped its hand on a new Codex feature called Scratchpad. It lets you fire multiple Codex agents in parallel from a TODO list view, with each task running asynchronously in its own isolated git worktree. The feature is still experimental, but it is almost certainly going to ship as part of the upcoming Codex Superapp that merges ChatGPT, the Atlas browser, and the coding tools into a single desktop product.

On its own, Scratchpad is a modest UI upgrade. Step back, though, and it is the clearest signal yet that parallel agent execution is no longer a feature. It is infrastructure.

What is Actually New

The mechanic is simple. You write a list of tasks in Markdown TODO style, hit go, and Codex spins up one agent per task. Each agent runs in its own git worktree, so they cannot stomp on each other, and you watch their status update live in a dashboard. When an agent is done, you review the diff and merge.

Cursor shipped something similar with its multi-pane parallel agents in Cursor 3. Claude Code has had worktree-based isolation for months. Windsurf's Arena Mode runs models side-by-side. What OpenAI is adding is the TODO-list-as-control-plane pattern. Planning and dispatch live in one flat view.

Why This Matters

Parallel agents change the economics of AI-assisted development in one specific way. They collapse the latency cost of waiting on a single long-running task. If each of five independent tasks takes 10 minutes, serial execution costs 50 minutes of wall time. Parallel execution costs 10. For any team where engineer time is the bottleneck, that is a real productivity lever.

Here is where I break from the hype. Parallel agents only help if your work actually parallelizes. Most real engineering work has dependencies. Task B waits on task A's interface. Task C needs to see task B's schema. Dumping five agents on a TODO list without thinking about the dependency graph produces five agents that either duplicate work or crank out conflicting changes that cost more to reconcile than running them serially would have cost in the first place.

My Take

The right mental model for Scratchpad is not "five agents instead of one." It is "an IDE for dispatching short, independent, well-specified units of work." Used that way, think five bug fixes across five unrelated files, it is genuinely useful. Used as a magic "ship my whole feature" button, it falls over fast.

The broader signal matters more than the feature itself. Every major coding tool is converging on the same primitives. Worktree isolation. Parallel dispatch. Long-running background agents. Model selection per task. If you are building a dev-tools product and you do not have these, you will not be competitive by summer.

If you are an engineer, the skill to actually invest in is learning how to decompose work into genuinely independent units. That compounds across every agent tool, not just OpenAI's.