Agent CLI game changers

Everyday The changes that actually alter how you work in Claude Code, Codex, Antigravity CLI and Grok Build. Not a full changelog.

Last checked: November 5, 2026.

CLI	Version checked
Claude Code	2.1.318
Codex CLI	0.166.0
Antigravity CLI	1.5.0
Grok Build	1.0.70

Commands were checked against each CLI's --help or official docs, but not all of them were run. These tools change weekly, so confirm on your own version.

The big picture

What changed	Why it matters
Gemini CLI gave way to Antigravity CLI	Google stopped serving Gemini CLI on personal plans from June 18, 2026
Agents run in the background	Start several, leave them running and check in from one dashboard. All four CLIs now have a version of this
You can move between tools	Claude Code, Codex and Antigravity can import setup from other agents, so switching costs less
Sandboxes became practical	Claude Code, Grok and Antigravity can run with OS-level limits, which beats approving every command
Old models retired	Check saved model names in your configs and scripts

Do this first

• Claude Code: if you use fast mode on Opus 4.8, move to Opus 5.5. It's removed on November 30, 2026, and requests will then fail with an error instead of falling back.
• Claude Code: check you're happy with auto mode. It became the default for new sessions on Pro, Max and Team plans from August 14.
• Claude Code scripts: claude ultrareview now needs workspace trust, so scripts must pass --trust-workspace.
• Codex: replace gpt-5.4, gpt-5.4-mini and GPT-5.5 model names, which were retired for ChatGPT sign-ins.
• Gemini CLI users: move to Antigravity with the migration steps.

Claude Code

Background sessions. Start a session, leave it running, and come back later:

claude --bg "fix the failing auth tests"
claude agents        # see and manage all sessions
claude attach NAME   # open one in this terminal
claude logs NAME     # recent output without attaching
claude stop NAME

Other changes that matter:

Feature	What it does
Auto mode by default	Fewer prompts on new sessions (Pro, Max, Team), with safety checks on risky commands
Forks and @ mentions	Fork mode is on by default, and @ refers to another session by name
/diff live panel	In fullscreen rendering, a panel beside the chat that refreshes as Claude edits
/skill-doctor	Shows what each skill costs in context and how often it's used
claude plugin eval	Tests a plugin against cases and compares it with no plugin
claude import	Imports config from another AI coding agent
--restricted	Starts without command-running tools or your user and project settings
permissions.blockWritesOutsideWorkingDirectories	Stops file tools writing outside your working folders, in every permission mode
Fable 5.1 and Opus 5.5	New models. Fable 5.1 has a 1M-token context window

Sources: What's new · Changelog

More here: Latest digest · Advanced

Codex CLI

An agent dashboard. Browse every session on your machine, then resume or fork one:

codex agents
codex resume --last
codex fork --last

Other changes that matter:

Feature	What it does
--approve-for-me	Automatic approval review for eligible requests, without widening file or network access
/import	Brings setup and recent chats over from Claude Code or Cursor
Worktrees (experimental)	Start or fork tasks in separate Git worktrees
@ task mentions	Reference other Codex tasks from the terminal
codex queue --resume	Queues a message for a session and resumes it if it's unloaded
codex doctor --network-only	Focused config, connectivity and WebSocket checks
GPT-6.1 Sol and Terra	New models. OpenAI suggests starting with Sol Medium for GPT-6 work, and Ultrafast mode speeds up Sol
model_provider = "gcp"	Built-in Google Cloud Vertex AI provider

Sources: Changelog · What's new · Releases

More here: Codex CLI

Antigravity CLI (agy)

The successor to Gemini CLI. Start with the Antigravity page.

Feature	What it does
agy plugin import gemini / claude	Brings your extensions and plugins across
--mode auto	Fewer prompts. Needs the terminal sandbox, and isn't available on business or enterprise accounts
/goal, /schedule	Run toward a goal, or on a timer or cron schedule
--remote-agent	Runs conversations on remote agents
/voice	Talk to it, or dictate
agy acp	Lets editors use it as their agent
PreInvocation hooks	Can stop a turn before the model is called, for example to block a prompt with sensitive data

Sources: CLI reference · Changelog

Grok Build

It's open source. xAI published the agent loop, tools, terminal UI and extension system under Apache 2.0 in July 2026. You can read what it does, or build it and point it at your own local models. Outside contributions aren't accepted.

A real deny list. Grok's sandbox can block paths at kernel level, including globs:

# .grok/sandbox.toml
[profiles.project]
extends = "workspace"
deny = ["**/.env", "**/*.pem"]

The sandbox is off by default, so turn it on with grok --sandbox workspace or [sandbox] profile. Built-in profiles don't permanently protect paths like ~/.ssh, so list those too.

Other changes that matter:

Feature	What it does
grok dashboard	Opens the Agent Dashboard view at startup
grok --sandbox profiles	workspace, devbox, read-only, strict
[memory] enabled	Cross-session memory. xAI's reference lists it as off by default
grok inspect	Shows which configs are being picked up
grok doctor	Checks terminal, clipboard, color and input support

xAI also announced Workflows in September 2026, but its changelog and news pages blocked automated reading, so check those directly: Changelog.

Check what gets uploaded

An independent analysis of version 0.2.93 reported repository content, including .env contents, being sent to xAI. I couldn't confirm it, or whether current versions behave this way, and I found no official statement. Until you've checked, keep real secrets out of repos you open with it.

Sources: Grok Build is open source · Sandbox · Settings · GitHub

More here: Grok Build · Advanced

Keep .env away from agents

Agents run commands as you, so a rule blocking .env may not stop a script from reading it. Controls differ:

CLI	Control	Note
Claude Code	Read(./.env) deny rule, plus the sandbox	Documented, but users have reported gaps
Codex CLI	Permission profiles	Check the docs for your version
Antigravity CLI	permissions allow, ask and deny rules	No .env-specific guidance found in the CLI reference
Grok Build	Sandbox deny list	Documented as kernel-enforced, once the sandbox is on

The dependable fix is the same everywhere: keep real secrets out of the project. Store op:// references in the file and inject values at run time:

op run --env-file .env.1p -- pnpm dev

More in Install script safety.

Related

• AI agent CLIs
• Latest digest
• Install script safety