Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Codex 0.167: run parallel sessions in one place

Written November 7, 2026

Codex 0.167.0 adds an agent command centre with unread-result markers and backend switching that keeps drafts. What it supervises — and what it still can't.

The problem it solves

Running several Codex agents in parallel means losing track of which sessions finished, which need input, and which drafts survive a backend switch. Codex 0.167.0, released November 6, 2026, adds a command centre for exactly that supervision. The release notes describe what it shows; they do not claim it decides anything for you.

What shipped

Open the agent command centre with codex agents --no-daemon, which runs it without a shared background server. The command centre and the session picker now mark unread results, and opening a viewed transcript acknowledges it as read, so a finished task stays distinguishable from one still running. Tasks in the command centre are sorted by recency within status groups, and a repository filter and directory grouping in the picker help navigate sessions across linked worktrees.

Switch between connected backends from the terminal session picker and keep your drafts when switching. Saved chats stay visible after you switch model providers; if a chat cannot resume, you can read its transcript and prepare an editable prompt to continue in a new chat. That is the practical change for multi-backend work: the draft and the history survive the switch, even when the session itself cannot.

Two companion changes ship in the same release. Preview local image attachments while composing on supported terminals, and open viewed local images from clickable transcript paths. And a gcp_setup_wizard sets up Google Cloud Vertex AI with Application Default Credentials or a credential file, with API key and cloud provider options under "Sign in another way." Patch details can also be expanded directly in the transcript while an approval prompt is open.

How to use the command centre

Start three or four real tasks in parallel — a bug fix, a test-writing pass, a refactor, a docs update — then open the command centre and work through the unread results first. The markers tell you what you have not looked at yet; they do not tell you which result matters. Read each finished result before starting the next round of work, so a session that finished with a question does not sit unanswered while you start something new.

When a task needs a different backend, switch from the session picker and keep the draft. The draft survives; the session history stays readable. If the chat cannot resume on the new backend, the transcript plus an editable prompt is the way back in — copy the prompt, not the whole transcript, and start the continuation as a new chat.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Run parallel subagents in Claude Code · Worktrunk: Git worktrees for parallel agents · VS Code agent bake-off: the same task, four agents