codex cli 0.167.0: Multi-Backend Agents Come to the Terminal

codex cli 0.167.0 is out, and it is the release where OpenAI's terminal coding agent stops behaving like a single-session tool. Published on GitHub on November 6, the release adds an agent command center, lets you switch between connected backends without losing drafts, previews local image attachments in the composer, and ships a Google Cloud Vertex AI setup wizard, alongside a set of reliability fixes that matter if you run the CLI all day. Dev Daily Briefs headlined its coverage "Codex CLI goes stable," and release trackers list 0.167.0 as the current stable line. Source

This post walks through what actually shipped, what the release notes themselves say, and where the practical upgrade decisions sit for teams running coding agents in production workflows.

The agent command center

The headline feature is the agent command center, opened with codex agents --no-daemon. The command center and the session picker now mark unread results, so an agent that finished while you were doing something else shows up flagged instead of silently complete. For anyone running more than one agent session at a time, this turns the terminal into something closer to a task queue: start work in one session, context-switch, and see at a glance which sessions have results waiting. Source

The --no-daemon flag is worth pausing on. Per the release notes, it opens the command center without running a background daemon, so the coordination view belongs to that invocation rather than a persistent service. The notes do not say what happens to coordination state once the session ends, so treat persistence as unverified and check it in your own setup if your workflow depends on it. Source

The unread-result markers are the detail most likely to change daily behavior. The release notes list them as new in this version: the command center and the session picker now surface which sessions have results you have not seen, and viewed transcripts are acknowledged. AI/TLDR's summary of the release describes the combination as "an agent command center and backend switching," which matches what the changelog shows. Source

Backend switching that keeps your drafts

The terminal session picker can now switch between connected backends, and your drafts survive the switch. The resume picker also paginates across backends and reuses connected backends, so moving between, say, a local session and a cloud-hosted one no longer means retyping the prompt you were composing. Source

Multi-backend is the strategic part of this release. A coding agent that can move between backends lets a team route work by cost, latency, or data-residency requirements without changing tools. The related sign-in changes point the same direction: API key and cloud provider options now live under "Sign in another way," and enabling gcp_setup_wizard walks you through Google Cloud Vertex AI setup with Application Default Credentials or a credential file. That is enterprise plumbing, and it suggests OpenAI expects the CLI to run against more than OpenAI's own hosted backend. Source

Two smaller quality-of-life additions round out the interactive work: local image attachments preview in the composer on supported terminals, with clickable transcript paths opening viewed images, and you can opt into macOS or Windows system proxy settings from /experimental. The proxy option matters inside corporate networks, where egress rules have historically been a common reason terminal agents fail on first run.

The quieter fixes that matter more

Releases earn their keep in the fixes, and 0.167.0 has several that address real daily friction:

• 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, instead of losing the thread entirely.

• Automatic context compaction now accounts for retained output from failed responses and responses with missing token usage data. Compaction bugs are subtle: when the compactor miscounts what is in context, the agent's later turns degrade in ways that look like model regression but are actually bookkeeping errors.

• Network approval prompts close when their requests end, and concurrent requests to the same endpoint prompt one at a time. Anyone who has watched approval prompts pile up during a tool-heavy run will recognize why this matters.

• Ctrl+L clears the terminal display while keeping earlier messages available through Ctrl+T, and long realtime V3 conversations save completed speech to chat history as the transcript grows.

• Linux musl builds use fewer allocator arenas, reducing idle memory after conversations unload.

All of these come straight from the release notes. Source

Governance signals in the changelog

The full changelog between 0.166.0 and 0.167.0 also references documentation work that AgentOps teams should note: the app-server README now covers managed network constraints under [permissions.network] and prompt-user command rules that require fresh human approval. The changelog also references several items around Guardian review of recorded exec calls and preserving tool-call evidence through that review. Source

We read these as directional signals rather than finished features, since they appear in the changelog and documentation entries rather than the headline feature list. The direction is consistent, though: approval flows that cannot be silently bypassed, network access as a declared permission, and review tooling for what agents actually executed. Those are the same control categories we wrote about in our async subagents post on background workers, where LangChain's async deep agents raised the same question of how a human stays in the loop when the agent runs unattended. The terminal is becoming a control surface, not just an input box.

Where this fits among agent runtimes

Codex CLI's move mirrors a broader pattern we have been tracking. Google's agent substrate, which hit 1.0 and joined the CNCF, standardizes the runtime layer for agents as we covered in our agent substrate post, while monitoring vendors like Honeycomb and PagerDuty are shipping agents that operate on production systems, covered in our sre agent post. Codex CLI 0.167.0 is the developer-facing end of the same trend: multi-agent coordination and backend portability arriving in the tool engineers already have open.

The practical difference between these layers is scope. A substrate handles orchestration across services. A monitoring agent watches production. A terminal command center coordinates the handful of agents one developer is steering right now. They are complementary, and a team building agent workflows will likely touch all three.

Should you upgrade

This section is our read, separate from the sources. For most users, yes: the fixes alone justify it, particularly the compaction accounting fix and the network approval prompt cleanup, since both remove failure modes that are hard to diagnose from the outside. The command center is additive; nothing in the release notes suggests existing single-session workflows change behavior.

Two caveats. First, the Vertex AI setup wizard is behind the gcp_setup_wizard flag, so treat it as early access and validate it in a test project before pointing real workloads at it. Second, if your team pins CLI versions in CI or dev containers, the changelog between 0.166.0 and 0.167.0 is long; skim it for anything touching your configuration before rolling the pin forward. The install is a single npm command, npm i -g @openai/codex@0.167.0, so testing in an isolated environment is cheap. Source

If you are evaluating coding agents more broadly rather than upgrading an existing install, the more useful frame is workflow fit: a terminal-native agent with backend switching suits teams that live in the shell and want provider flexibility.