No new model, but a command center: Codex CLI 0.167.0 learns to supervise agent fleets
OpenAI released Codex CLI 0.167.0 on November 6: no new model, but an agent command center, unread markers, and drafts that survive backend switches. The coding-agent race is shifting from how smart the model is to whether a human can manage a fleet of them.

An update with no new model — and why it still deserves its own article
At 01:40 UTC on November 6, 2026, OpenAI released Codex CLI 0.167.0, tagged rust-v0.167.0. It is a stable release with no new model in it. The release notes list 136 changes, and the first item is not a performance number but a command: codex agents --no-daemon. A first glance might disappoint: no new model, no price cut, no benchmark scores. But that is exactly why it is worth writing about — once coding agents go from "one assistant" to "a team running several jobs at once," the bottleneck is no longer the model. It is how a human keeps track of all of them. Version 0.167.0 is an update about managing people and tasks.
Some background first. Codex CLI is OpenAI's open-source terminal coding agent; the code lives in the openai/codex repository on GitHub under the Apache-2.0 license. It ships fast: five stable releases between October 15 and November 6, 2026, with 0.166.0 on October 29 and 0.167.0 just eight days later. VibeFix covered versions 0.161 and 0.164 before; 0.167.0 is not a routine patch but the release that fills in a missing piece for running multiple agents at once.
The command center: seeing every task at a glance, at last
The headline feature is the agent command center, opened with codex agents --no-daemon. It shows every agent task in one view — running, queued, and finished — grouped by status and sorted by recency within each group. The command center and the session picker now mark unread results, so you can tell at a glance which jobs produced something you have not looked at yet; opening a transcript marks it as viewed. That sounds small, but anyone who has run agents in parallel knows its weight: before, you had to flip through sessions one by one to confirm what finished; now the system tells you "there is something new here."
One common misunderstanding is worth clearing up: --no-daemon does not turn off the background service. Since 0.157.0, interactive sessions start a shared background server automatically (daemon_auto_start), and that default has not changed. The flag only opens the command-center view. OpenAI did not touch the background architecture; it just added a way to see the whole picture without going through the daemon. Use the flag if you want it; nothing changes if you do not. There is no migration cost.
The command center also works for embedded sessions: the same release enables it there (pull requests #58238 and #58239) and opens task tools to embedded app-server sessions. That matters less to terminal users, but it signals that OpenAI treats the command center as a platform-level capability, not a small panel inside the terminal. One more detail: tasks in the command center are grouped by status and sorted by recency within each group (#58714). Sorting sounds trivial, but it decides what you see first when you open the view — the jobs with recent activity, not an alphabetical dead list. For someone running several jobs a day, seeing the right thing first matters more than another filter.
The shape of 136 changes: features on top, control underneath
The release notes list 136 changes, but the distribution is revealing. There are only 7 new features, 9 bug fixes, 2 documentation items, and 2 chores; the remaining 100-plus entries are the full changelog, listed by pull-request number. Judging by the feature list alone, this looks like a small release. Read the changelog and a different picture emerges: a large share of the work is about Guardian, Codex's built-in review mechanism — direct review of recorded exec calls, experimental review of injected thread commands, elastic budgeting and caps for Guardian context, shared token limits for Guardian context delivery, and the ordering of PreToolUse hooks before Guardian review (PRs #58028, #58119, #58282, #58444, #58454, #58368, among others).
In other words, where OpenAI really spent effort in this release was not making agents more capable but making them more controllable. The documentation changes point the same way: the app-server README now explains managed network constraints and command rules requiring fresh human approval, and the openai-docs skill recommends the documented replacement when migrating away from a deprecated model. Codex is moving from a personal productivity tool toward development infrastructure that can run compliantly in a managed environment. That means little to an individual user, but for teams trying to roll Codex out inside a company, these changes matter more than the command center — a tool that fails security review never gets adopted, however good it is.
The changelog also carries hints of what comes next. Server-side truncation for Code Mode sessions, bounded output collection, and evidence chains for native tool calls do not have visible effects yet in 0.167.0, but they are groundwork for the Guardian v3 evidence pipeline, which third-party trackers say lands in 0.168.0. So 0.167.0 is a foundation-laying release: a command center on the surface, and stricter review and auditing being built underneath. Knowing that, the 136 changes in a release with no new model stop looking strange.
Switching backends without losing your draft
The second group of changes fixes another real pain point: switching backends. Codex can talk to different model providers — OpenAI's own, Google Cloud Vertex AI, and others. Switching providers used to break things: saved chats could vanish from the resume list, or refuse to reopen over a provider mismatch, a missing configuration, or a policy conflict. Work, in effect, was lost. Version 0.167.0 fixes that: the terminal session picker can now switch between connected backends and keep your drafts, and the resume picker, lookup by name, and resume --last now list chats from every provider, not just the current one.
And if a chat genuinely cannot resume, 0.167.0 offers a way back: you can read its full transcript and prepare an editable prompt to continue in a new chat. It is not a glamorous feature, but it is the line between a productivity tool and a toy — a productivity tool cannot afford to lose your work in progress. A related fix: automatic context compaction now accounts for retained output from failed responses and responses with missing token usage data, so the books balance after compaction.
The same batch adds a setup wizard for Vertex AI. Version 0.165.0 added a built-in provider for Google Cloud's Vertex AI; 0.167.0 adds an optional guided setup (behind the gcp_setup_wizard feature flag) using Application Default Credentials or a credential file. Sign-in options were tidied too: API key and cloud provider options now sit under "Sign in another way." The direction is clear: Codex is no longer built only around OpenAI's own accounts; it is becoming a terminal that can plug into several clouds.
Images, approvals, and proxies: small changes aimed at fewer interruptions
The remaining new features are small, but they share one goal: fewer interruptions while you supervise agents. Local image attachments can now be previewed while composing on supported terminals, and viewed images open from clickable transcript paths — before, attaching an image was mostly guesswork until the model complained. While an approval prompt is open, patch details can be expanded directly in the transcript instead of diffing elsewhere. System proxy settings on macOS or Windows can be opted into from /experimental. Network approval prompts now close when their requests end, and concurrent requests to the same endpoint prompt one at a time instead of piling up a row of confirmation boxes.
Two easily missed fixes: pressing Ctrl+L to clear the screen no longer loses earlier messages — they remain available through Ctrl+T — and long realtime V3 conversations now save completed speech into chat history as the transcript grows, instead of holding it until the end. None of these are headline features, but they fix the small snags that make a tool feel unfinished. For a tool shipping weekly, keeping users often comes down to exactly this kind of repair.
There is movement on the documentation side too: the app-server README now covers managed network constraints under [permissions.network] and prompt-user command rules that require fresh human approval, and the openai-docs skill recommends the documented replacement when migrating away from a deprecated model. Read together, OpenAI is moving Codex from "it runs" toward "it runs compliantly in a managed environment." Much of the release — the Guardian review of recorded exec calls, experimental review of injected thread commands, elastic budgeting for Guardian context — is about keeping agents under control rather than making them more capable.
In context: three vendors patching the same gap
Put 0.167.0 on the early-November timeline and it stops looking isolated. The same week, Claude Code was tightening auto mode and permission checks, GitHub Copilot was spreading into Xcode and JetBrains Air, and Devin was shipping a cheaper model in SWE-2 Flash. New-model announcements were slowing while "how do you manage several agents at once" tooling was speeding up everywhere. That is not a coincidence: when one agent can work for eight days and coordinate 1,000 subagents — Anthropic's Fable 5.2 result, published the same week — the human job shifts from writing prompts to being the team lead: assigning work, checking results, cleaning up messes. Whoever makes that job pleasant first wins the next wave of users.
From that angle, 0.167.0's command center arrives right on time, but it is only a start. Today's version shows status and marks unread items; it cannot yet batch-operate, group by project, or be checked from a phone. OpenAI left its own tell in the notes: the 0.168.0 alpha series is already running, and Guardian v3 evidence-pipeline work lands in the next stable release. The race to build supervision tooling has just started; 0.167.0 is OpenAI's first entry.
Before you upgrade
Upgrading is simple: npm install -g @openai/codex@0.167.0, or wait for your package manager. Besides npm, Codex ships via Homebrew, direct binaries on GitHub Releases (macOS, Linux, Windows), and DotSlash manifests for pinning a version in shared team configuration. Version 0.167.0 is stable, not a prerelease, so it is safe to install directly. Two things are worth knowing. First, the release cadence is fast — five stable releases between October 15 and November 6, with alphas landing almost daily — so if you pin versions in automation scripts, keep them current rather than letting them rot on an old release. Second, --no-daemon is only a view switch; it does not change the background service's default behavior, so if you had concerns about the daemon, this release does not address them. Do not misunderstand it.
One more caution: the release notes contain no benchmark scores or productivity figures. Every change describes behavior, not measured results. How much time the command center actually saves you in flipping between sessions is something only your own use can answer. We will not claim a big efficiency gain from the release notes alone, and neither should you. Two related items from the same week are worth mentioning with their credibility kept separate: OpenAI announced on November 4 that it had agreed to acquire Pierre Computer Company, subject to regulatory approval — that is official. Reports of a GPT-5.3-Codex model, by contrast, rest so far on release trackers with no official announcement to back them, so this article does not treat them as fact. In a week this dense with news, telling "in the official notes" apart from "reported by a tracker" is basic hygiene.
Verdict: the first half of the model race is over; the second half is about staying in control
Codex CLI 0.167.0 is easy to dismiss as a boring update: no new model, no price cut, no record-breaking score. But it marks a shift: the competition in coding agents is moving from "how smart is the model" to "can a human manage a group of models." A command center, unread markers, drafts that survive backend switches, chats that stay recoverable — none of these belongs on a keynote slide, but each one answers a real frustration from daily use. OpenAI chose this moment to fix them in a batch, which suggests it has heard the same complaint repeatedly: the models are strong enough now; the hard part is running five jobs at once without chaos.
My judgment: if you already run parallel tasks in Codex, upgrade now — the command center and the cross-backend fixes map directly onto daily pain. If you only use Codex for single tasks in a single session, this release will barely register, and waiting for the next one is fine. For the industry, 0.167.0 reads more like a signal: in November 2026, the coding-agent arms race entered its second half, and it will be won not by whoever has the highest benchmark score, but by whoever lets one person direct ten agents without losing control.
From tool to console: leave a supervisor's seat in your vibe project
The most interesting change in 0.167.0 is not any single feature but a shift of role: Codex is no longer just a terminal you ask and it answers — it is starting to behave like a task console. That shift matters directly for vibe coding workflows. Before, one agent did one job at a time, and you watched it finish before starting the next. Now you can run several jobs at once — one writing a feature, one adding tests, one checking the docs — and the command center is the chair you sit in to watch the whole board. The chair writes no code, but it decides whether you can manage three jobs at once without losing track.
One detail is easy to miss: drafts that survive a backend switch. Switching backends used to mean throwing away a half-finished job and starting over, which kept many people from ever switching. Now the draft travels with you, so the job itself has continuity, and you can pick the cheaper, faster, or steadier backend without paying for the switch with rework. For an independent developer, that detail is more practical than the command center itself: it turns switching backends from a gamble into an everyday choice.
But it is worth saying plainly what the command center cannot do. It can show you the state of five jobs, but it cannot judge which one has gone off track; it can flag unread updates, but it cannot read them for you and tell you what actually changed. Supervision is still a human job — glance, judge, stop. If you never had the habit of checking what your agents are doing, a command center will only let you ignore five jobs at once instead of one. The tool has been upgraded; the habit has to follow, or the console is just a prettier dashboard.
One final, blunt point: terminal tools are not decided by release notes but by how often you open them each day. Whether the command center is good depends not on how many features it lists, but on whether you still type codex agents a week from now. If you already juggle three or more jobs a day, install 0.167.0, run it for a week, and then judge whether it earns its keep. If you still work one task at a time, there is no rush — the tool will wait until the day you need it.
Comments (0)
Sources
Related articles
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.