Claude Code makes Auto mode the default today for Pro, Max, Team
A classifier that blocked 89% of dangerous commands now runs your approvals — plus Codex reaches Linux and Zed opens Delta for agents.
Copy markdown
On by default, starting today
New Claude Code sessions on Pro, Max, and Team plans now launch in Auto mode as of today — Claude runs tools without stopping for a Yes/No on each call. Auto mode has existed since March; what flips today is that it's the default, not opt-in.
The classifier that replaces your Yes/No
Instead of prompting you, a classifier routes every tool call and blocks anything irreversible, destructive, or aimed outside your environment. Anthropic says it caught 89% of dangerous commands across 1,053 test scenarios, versus 13.6% for human reviewers — whose attention drops off past ~50 prompts.
You still hold the kill switch
Three consecutive blocks — or twenty in a session — drop you back to manual approvals, hard-deny rules override everything, and prompt-injection screening guards against exfiltration. Enterprise, the API, Bedrock, Vertex, and Foundry stay opt-in, and you can switch Auto mode off entirely.
Elsewhere: Codex desktop reaches Linux
OpenAI shipped its ChatGPT + Codex desktop app to Linux in preview on Aug 13 — .deb and .rpm builds for Ubuntu 24.04/26.04, Debian 13, and Fedora 43/44 on x64 and ARM64. Codex now works against your local files without a browser tab, closing OpenAI's last desktop gap.
Elsewhere: Zed opens Delta for agents
Zed launched Delta, a multiplayer coding environment where humans and AI agents work the same project in real time, paired with DeltaDB's history that records every edit and its agent context. If you build collaborative agent tooling, it's a weekend worth exploring.