AI Coding Agent Latest News 2026: What Shipped from 7 to 14 November

The latest AI coding agent news for 2026, for the week of 7 to 14 November: Claude Code added an experimental Code Mode and plugin version ranges, OpenAI's Codex CLI shipped side-chat tabs and stricter handling of untrusted folders, GitHub Copilot CLI shipped five stable releases, OpenCode 2.3 made subagents run in the background by default, goose added deferred MCP tools, and Gemini CLI 0.68 was mostly fixes.

Every item below comes from the project's own release page on GitHub, and the dates are the GitHub publish dates in UTC, as of 2026-11-14. Nothing here was installed or run for this post except the one command in the how-to section, which was run on 2026-11-14.

The week in one table

Agent	Stable releases, 7 to 14 Nov 2026	Headline change	What it means for MCP and skill users
Claude Code	2.1.321 (9 Nov), 2.1.322 (10 Nov), 2.1.323 (12 Nov), 2.1.325 (13 Nov)	Experimental Code Mode; plugins can state the Claude Code versions they support	MCP OAuth redirect fix in 2.1.321; plugin authors can pin a version range
Codex CLI	0.167.2 (11 Nov), 0.168.0 and 0.169.0 (both 13 Nov)	Side conversations, then multiple side-chat tabs	codex exec keeps project MCP servers off in untrusted folders
GitHub Copilot CLI	1.0.108 (9 Nov), 1.0.109 (10 Nov), 1.0.110 (12 Nov), 1.0.111 and 1.0.112 (13 Nov)	Masked dialog for git, ssh and sudo password prompts (experimental)	Skill picker no longer matches words hidden in escape sequences; ACP hook fix
OpenCode	2.2.7 (10 Nov), 2.3.0 (13 Nov)	Subagents run in the background by default	Experimental MCP events; remote MCP logins stay signed in
goose	1.58.0 (9 Nov), 1.58.1 and 1.59.0 (10 Nov), 1.59.1 (14 Nov)	Deferred MCP tools; context window report	Fewer tool definitions loaded up front
Gemini CLI	0.68.0 (10 Nov)	Mostly fixes; new extension validation command	Extension authors can validate before publishing

Pre-releases (Copilot CLI's -0, -1 builds and Gemini CLI 0.69.0-preview.0) are left out of the table.

Claude Code: Code Mode and plugin version ranges

Claude Code 2.1.323, published 12 November, added Code Mode, described in the release note as "an experimental opt-in in which Claude writes JavaScript to call its tools". The Code Mode docs cover turning it on, and turning it off for an organization.

The same release added engines["claude-code"] to plugin.json. Plugins outside their stated Claude Code version range are not installed, updated or loaded, and install picks the newest plugin version that fits. If you ship a plugin or skill bundle, this is the field to set before your next release. claude plugin validate also now type-checks a plugin's TypeScript hooks module, with --strict turning warnings into failures.

Two fixes are worth knowing if you run MCP servers:

• 2.1.321 (9 November) fixed MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them. It also fixed a case where a read-only Bash command ran outside the sandbox without a permission prompt.
• 2.1.325 (13 November) fixed "a security issue in how zstd-compressed responses are decoded" and fixed subagents on a different model sizing MCP results with the main session's model.

2.1.321 also added Claude Fable 5.2 (claude-fable-5-2) as the default Fable model, listed at $8 input and $40 output per million tokens with 1M context, per that release note. 2.1.322 added fast mode on Vertex in the global, us and eu regions.

Codex CLI: side chats and stricter untrusted folders

Codex CLI 0.168.0 added a side conversation next to the main chat behind features.tui_workspace, and a model provider browser. A few hours later 0.169.0 let that opt-in workspace hold multiple side-chat tabs and added task search in the agent command center.

The change to read twice is in 0.169.0's bug fixes: codex exec now keeps project configuration and MCP servers disabled in untrusted directories, even with --sandbox workspace-write. To save trust explicitly, the note says to pass --trust to codex exec or codex review. If a CI job relied on a repo's own MCP servers loading under codex exec, check it after upgrading.

0.168.0 also warns when features.multi_agent is set and points to agents.enabled for turning off subagent tools. 0.167.2 made incremental tool updates the default on Google Cloud Vertex AI.

GitHub Copilot CLI: five stable releases in five days

Copilot CLI published 1.0.108 through 1.0.112 between 9 and 13 November. Most lines are fixes. The ones that change how you work:

• 1.0.108: with the sandbox on, a running local MCP server restarts after /cd so it follows the new directory's denied paths. copilot help config now names the .claude/rules instruction files and .claude/settings.json that the CLI reads. Skill enable and disable lists now match names in any letter case.
• 1.0.109: IPv4 and IPv6 CIDR ranges in sandbox host allow and deny rules, Claude Fable 5.2 in the model picker, and in ACP mode a preToolUse hook that answers "ask" now asks the ACP client instead of denying the call.
• 1.0.110: an experimental masked dialog for git, ssh and sudo password prompts from agent commands. The / picker no longer runs a skill for a word its description hides inside a terminal escape sequence.
• 1.0.111: regular expressions in enterprise-managed Shell, Bash and PowerShell permission rules, and no duplicate MCP OAuth reconnects.

The skill picker fix is a small one with a clear lesson for skill authors: a skill's description is input the tool reads, and text a person cannot see should never decide which skill runs. Our agent skills guide covers how descriptions are used.

OpenCode 2.3: background subagents by default

OpenCode 2.3.0 (13 November) made subagents return immediately, keep running in the background and report back when idle. Transcripts now collapse each run of tool calls into one row; set Verbosity to medium or high to see every step. It also added experimental MCP events behind the mcp_events flag and client certificates for gateways that require mutual TLS.

2.2.7 (10 November) kept remote MCP logins signed in when servers rotate refresh tokens, and has breaking changes for code that imports @opencode/ai directly: ImageResponse.image is removed.

goose and Gemini CLI

goose 1.59.0 (10 November) added minimal deferred MCP tools, session filtering by client tag for ACP clients, and resuming a finished subagent with a follow-up. 1.58.0 added a context window report and viewer. The goose repository now lives under the aaif-goose GitHub organization.

Gemini CLI 0.68.0 (10 November) is mostly fixes, including case-insensitive root checks on Windows, plus a new extension validation command.

The pattern this week

Three things showed up in more than one tool, each in its own release notes:

1. Subagents run more on their own. OpenCode made background subagents the default, goose can resume a finished one, and Claude Code and Codex both fixed subagent model and setting handling.
2. MCP login handling got fixed in three places. Claude Code (OAuth redirects), OpenCode (rotating refresh tokens) and Copilot CLI (duplicate reconnects) all shipped MCP OAuth fixes.
3. Plugins and extensions get checked earlier. Claude Code added version ranges and stricter plugin validate, and Gemini CLI added an extension validation command.

If you build an MCP server, the practical takeaway is to test your OAuth flow against the new releases, and read our MCP server security guide for what a client can and cannot protect you from.

How to check what shipped yourself

You do not need a news site to see what shipped. GitHub's public releases API lists every release with its publish date. This command lists stable Codex CLI releases since 7 November; it was run on 2026-11-14 and printed rust-v0.169.0, rust-v0.168.0 and rust-v0.167.2:

curl -s "https://api.github.com/repos/openai/codex/releases?per_page=10" | python -c "import json,sys; [print(r['published_at'][:10], r['tag_name']) for r in json.load(sys.stdin) if not r['prerelease'] and r['published_at'] >= '2026-11-07']"

Swap the repository for anthropics/claude-code, github/copilot-cli, google-gemini/gemini-cli, anomalyco/opencode or aaif-goose/goose, and change the date to the start of the week you want. Unauthenticated calls to the GitHub API are rate limited, so run it a handful of times, not in a loop.

Before you upgrade a coding agent, read the note for three things: breaking changes, anything that touches MCP or OAuth, and anything that changes what the agent may run without asking. When you pick tools to give that agent, check the "What it can touch" line on each listing in our agents directory.

FAQ

Where can I find AI coding agent latest news in 2026?

The most reliable source is each project's release page, because it is dated and written by the maintainers. For Claude Code that is the changelog and GitHub releases; for Codex it is the Codex changelog and the openai/codex releases. Our news page collects agent news as it lands.

Where do I find each AI coding agent new release from September 2026?

Use the command above with the date set to 2026-09-01 and a larger per_page value, or open the project's releases page and scroll to September. This post covers 7 to 14 November only, so it does not list September releases.

Is there a summary of AI coding agent news from October 2026?

Not in this post, which covers 7 to 14 November only. The same release pages hold every October release with its date, and Claude Code's changelog and the Codex changelog are the quickest to read month by month.

Which change this week matters most for MCP server authors?

The MCP OAuth fixes in Claude Code 2.1.321, OpenCode 2.2.7 and Copilot CLI 1.0.111, plus Codex 0.169.0 keeping project MCP servers off in untrusted folders under codex exec. If your server uses OAuth with refresh token rotation, test sign-in again against these versions.

List your agent or MCP server

Built a coding agent, MCP server or skill? List it free on AgenticMart. Each listing shows what it can touch. List free now, paid selling opens soon.