What changed in coding agents in October 2026

Fast mode for Opus 4.8 in Claude Code stops working on 30 November 2026. Per the 2.1.316 release notes, requests that ask for it after that date return an error and do not fall back to standard Opus 4.8. Anthropic’s advice is to move to Opus 5.5, and that’s the change from October most likely to break something unattended.

The other change worth acting on is a new Claude Code setting that stops its file tools writing outside the project. The rest of the month’s single-tool changes are below, grouped by tool, each with the link to read before you rely on it.

Claude Code

Version 2.1.316 (31 October) deprecates Opus 4.8 fast mode as of 30 October and removes it on 30 November. Anything that sets Opus 4.8 with fast mode on, whether a settings file, a script that launches claude -p, or a CI job, needs a new model name before then. Search for it now, because the release notes describe an error on the request, not a fallback, which means an unattended job fails partway through instead of running a little slower. A nightly script that nobody watches is the likeliest casualty, since the first person to notice is whoever reads the missing output the next morning. The Opus 5.5 pricing post covers what the move costs.

Version 2.1.314 (30 October) adds permissions.blockWritesOutsideWorkingDirectories. With it set, Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode, including the ones that otherwise skip prompts. It is the write-side partner of permissions.blockReadsOutsideWorkingDirectories, which the settings docs describe as honoured when set to true from any scope. The changelog line doesn’t give the value format for the new key, so check the settings reference before committing it to a shared .claude/settings.json. Note what it covers: the changelog names the file tools, and a shell command that redirects output with > is a different path, so keep your Bash permission rules and sandbox in place alongside it.

Two smaller items in the same stretch. Version 2.1.316 adds CLAUDE_CODE_DISABLE_ULTRATHINK, which turns the ultrathink keyword off, and 2.1.311 (26 October) changed ultrathink to run at xhigh effort on several models. If a teammate pastes the word into prompts out of habit, that is now a real cost decision, and the variable lets a team switch it off.

Codex

Codex CLI 0.164.0 (23 October) adds ten numbered /presets slots for saving reusable prompts and slash commands, with a keyboard menu. It also lets /model and /permissions take inline arguments, so a model, reasoning effort, speed and permission profile can be set in one line instead of through a picker. Worth it if the same three setups come up daily (a cheap model for exploration, a stronger one for review, a read-only profile for unfamiliar code).

Codex CLI 0.165.0 (27 October) adds a built-in Google Cloud Vertex AI provider. It is configured in config.toml:

model_provider = "gcp"

That matters for teams whose Google Cloud billing and data-residency rules already cover Vertex and who don’t want a separate vendor account. The same release turns memory v2 on by default for fresh profiles and adds importing memories from other agents. Existing profiles aren’t switched over, so the change shows up on a new machine or a new hire’s laptop first.

The three Codex releases fall within one week, so a team on a pinned version may have skipped all of them. Codex CLI 0.166.0 (29 October) adds codex queue --resume, which resumes unloaded threads after queueing a message, and codex doctor --network-only for focused configuration, connectivity and WebSocket diagnostics. The second one is the useful one: when Codex fails behind a corporate proxy, run it first and attach the output to the support ticket.

codex doctor --network-only

Cursor

Cursor’s Agent SDK arrived on 20 October. An agent is defined as a code folder, tested with evals, and deployed either to Cursor’s infrastructure with agent-sdk deploy or to your own. For a team that already has a repeatable chore (a triage bot, a release-notes writer), the useful part is the evals: the agent’s behaviour gets a test suite that runs before each deploy. The changelog entry is short on detail, so treat it as a place to start reading, and judge the SDK on whether your first agent’s evals catch a regression you introduce on purpose.

Gemini CLI

Gemini CLI v0.66.0 became the stable release on 27 October. Most of it is terminal polish, but a few changes touch automation. Settings placeholders now resolve after .env files load, so a value like a project ID can come from the file. Interactive credential and SSH prompts are disabled in git operations, which stops a headless run hanging on a password prompt. Parsing of SKILL.md frontmatter was hardened for single-line descriptions. The release also carries a banner saying the CLI is not yet optimised for Gemini 4, so if you’ve switched models, expect rough edges. The Gemini 4 cost check has the numbers.

GitHub

Agent Merge, previously in the Copilot app and VS Code, reached pull requests on GitHub on 28 October. Choose Watch and fix in the merge box and Copilot starts work in a cloud development environment when a selected blocker appears: a failing Actions workflow, a merge conflict, or both. It combines with auto-merge, and a paid Copilot plan is required, subject to organisation policy. The changelog names only those two blockers, so don’t expect it to chase review approvals. Try it on a pull request with a flaky-test failure before turning it on by default, and read the commits it pushes as you would anyone’s.

Agentic autofix for code scanning alerts went generally available on 15 October. Copilot proposes a fix, reruns CodeQL to validate it and opens a draft pull request. Two limits decide whether it’s for you: it works only on alerts on the default branch, and it needs Code Security or Advanced Security on GitHub Enterprise Cloud plus a Copilot billing method. Usage draws on AI Credits and Actions minutes, and one pull request can cover up to 25 selected alerts. Batch the alerts by rule so each pull request is reviewable.

Two earlier items from October already have their own posts: assisted permissions and local sandboxing , both generally available.

Try this week

Grep your repositories, CI configuration and shell profiles for Opus 4.8, then check each hit for fast mode. Every job that requests it errors on 30 November, and it takes five minutes to find them now. The pattern below is deliberately loose, so expect a few false positives to dismiss by eye.

git grep -n -i -E "opus.?4.?8" -- . ':!node_modules'