A 1,053-person study found humans caught a planted dangerous command just 13.6% of the time (5% after 50 prompts from approval fatigue). Auto mode blocked the same commands 89% of the time, flat across session length.
Starting August 14, Auto mode is the default permission mode for Pro, Max, and Team plans. A separate classifier reviews every shell command and tool call; in testing it caught 89% of dangerous commands versus 14% for manual approval.
The classifier that powers Auto mode runs on every tool call and uses a small number of extra tokens. Those tokens no longer count against usage limits on Pro, Max, and Team plans, effective immediately. Enterprise and API users will get the same treatment once Auto mode becomes their default.
Boris Cherny says the entire team has been running Auto mode exclusively for months and couldn't imagine returning to manual permission prompts, ahead of Auto mode becoming the default on August 14.
Sessions can now pass context to each other without the user re-explaining anything. Claude sends a summary (not full history or files) so the receiving session can pick up mid-task.
Sessions can query each other bidirectionally: you can ask another session a question and get the answer piped back into your active session. Claude can also initiate these messages autonomously, for example when a change in one session affects work happening in another.
Two tuning improvements landed for Claude Tag: unprompted responses are down 30%, and Sonnet 5 posting in channels instead of threads is down 90%. The team is still actively collecting feedback.
After users reported that Claude Tag kept ignoring preferences for threaded comment replies, an Anthropic team member confirmed improvements have been shipped and will roll out to new sessions gradually. Users are encouraged to keep submitting feedback.
Hooks are more reliable than system prompts for controlling behavior in context-heavy Slack environments. Smaller models tend to forget general prompting instructions, so interrupting with a hook at the moment they attempt a top-level post and asking them to confirm intent works better.
The team's advice for getting the most out of Opus 5: give it context and goals and let it find its own path. If you want to be specific, collaborate on a detailed plan via artifacts first, then hand off implementation.
A user reported that Opus 5 makes worse assumptions and loses context compared to an earlier model. The team asked for specific transcripts via /feedback so they could investigate.
Auto mode pairs especially well with Opus 5 for long-horizon work: the model can plan and execute multi-hour tasks without stopping at every tool call for a permission prompt. Users can kick off parallel tasks and return to completed PRs.
In a study with 1,053 paid testers, humans caught a clearly dangerous command 13.6% of the time, dropping to about 5% after 50 prompts due to approval fatigue. Auto mode caught the same commands 89% of the time, with no degradation over session length.
When the August 14 default flip happens, users will get an in-app notification. Anyone who has already set a permission mode will be prompted before anything changes. Modes can still be toggled with Shift+Tab, and admins can pin or disable Auto mode in managed settings.
