Claude Code Is Making Auto Mode the Default, and It Now Talks Between Sessions
Quick verdict
Anthropic is changing how Claude Code handles permissions. Instead of asking you to approve every shell command, auto mode runs a separate classifier that reviews each action before it executes, and it will become the default for Pro, Max, and Team users. Anthropic says the classifier caught 89% of dangerous commands in testing, against 14% for manual approval on its own. Alongside that, Claude Code now lets one session message another, so a coding agent on one machine can hand context to an agent somewhere else without you shipping files around.
What actually shipped
Two changes landed together, and both point at the same thing: running more than one agent at once is becoming the normal way to use Claude Code.
The first is auto mode as the default permission path. Today most people run Claude Code in a mode where it stops and asks before touching your shell. That is safe, but it also means you sit there approving npm install forty times a day. Auto mode replaces the constant prompts with a classifier that looks at each command and action, decides whether it is risky, and only escalates to you when something looks dangerous. In Anthropic's own numbers, that classifier flagged 89% of dangerous commands, while manual approval alone caught 14%. The gap is the interesting part. People click "yes" on approval dialogs without reading them, so a reviewer that never gets bored beats a human who does.
The second is cross-session messaging. One Claude session can now summarize its work to another session on any machine, rather than transferring full files and chat history. If you have a planning session and a separate build session, the planner can pass a short brief to the builder instead of you copying and pasting a wall of context. Anthropic paired this with a few smaller managed-agent updates: session budgets that cap how much a run can spend, automatic loading of the skills stored in a repo, and "advisor" models you can call mid-session for a second opinion.
| Change | Before | After |
|---|---|---|
| Permissions | Manual approval per command | Classifier-reviewed auto mode by default (Pro/Max/Team) |
| Dangerous-command catch rate | 14% (manual) | 89% (classifier, per Anthropic) |
| Sharing context | Copy files and history by hand | Session-to-session messaging |
| Cost control | Watch the token meter yourself | Session budgets |
Why it matters
The permission change is a real UX bet, and it cuts both ways. If you have ever left Claude Code half-finished because you got tired of babysitting approvals, auto mode removes that friction. The flip side is that you are now trusting a classifier to catch the rm -rf that a tired human would have caught. Anthropic's 89% figure is its own testing, not an independent audit, so treat it as a strong internal claim rather than a settled fact. If you work in a repo where a wrong command is expensive, keep an eye on what auto mode lets through for the first week.
Cross-session messaging is the quieter but arguably bigger shift. It turns "run a coding agent" into "run a small team of coding agents that talk to each other." That is the same pattern people have been hacking together with shared files and scratchpads, now built into the tool. For anyone comparing coding agents, this moves the goalposts: the question is no longer which single agent is smartest, but which one orchestrates a group of sessions with the least babysitting.
There is a cost angle too. Session budgets exist because multi-agent runs burn tokens fast, and Anthropic clearly knows it. If you are already juggling several paid AI tools, giving one agent the keys to spin up more agents is exactly the kind of spend that creeps up on you. Capping it per session is a sensible default.
Video: what changed in Claude Code
A walkthrough of the recent Claude Code changes, including how the agent view and permission handling now behave.
FAQ
Do I have to use auto mode?
No. Auto mode becomes the default for Pro, Max, and Team users, but the manual approval mode still exists. If you want to review every command, you can switch back.
Is auto mode safe?
Anthropic says its classifier caught 89% of dangerous commands in testing, versus 14% for manual approval. That is its own internal number, so watch what it allows on your own repos before trusting it blindly. For high-stakes work, keep a close eye on the first few sessions.
What does cross-session messaging actually do?
It lets one Claude Code session send a summary to another session, even on a different machine, instead of you copying files and chat history between them. It is aimed at multi-agent workflows where a planner hands work to a builder.
Will running multiple agents cost more?
Yes, multi-agent runs use more tokens. Anthropic added session budgets to cap spend per run. If you pay for several AI tools already, see our Claude Code pricing breakdown before scaling up.
Sources
- @ClaudeDevs - cross-session messaging in Claude Code
- @ClaudeDevs - auto mode becoming the default permission mode
- @ClaudeDevs - blog post on the auto mode classifier
- @ClaudeDevs - session budgets, repo skills, and advisor models
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix