Claude Code Just Got Scary Good at Autonomous Workflows

[00:00:00]Claude's code is being released faster than anyone can keep up with it. Among this flow, there are four things that are changing your work. A fleet of agents that drives itself. A branching command for your conversations. Claude inside Chrome for everyone. And a million context tokens.

[00:00:17]Here is the main news.

[00:00:18]Your agents create pull requests themselves. And now a change that will put an end to constant surveillance. I left it for last.

[00:00:25]Six changes deserve a place on this board.

[00:00:28]First, sub-agents now run in the background by default, and when they finish working on the code, they themselves commit, push, and create a draft merge request. Second, branch your conversation into a background session and continue working where you were. Third, Claude in Chrome has exited preview for everyone.

[00:00:47]Fourth, 3.5 is a new default model with a window of 1 million tokens. Fifth, screen reader mode makes the entire terminal accessible. And sixth, streaming now consumes 37% less CPU resources, and the binary file is 7 megabytes lighter. The fleet is a true story.

[00:01:06]Run subagents and they will no longer block your session. They work in the background while you continue typing. Up to 200 of them can run simultaneously before the limit is triggered.

[00:01:17]Notifications are sent when an agent completes work or requires data entry. The Agents view shows the entire list, and Ctrl + X clears completed work trees without losing unsent comments. Background agents even update themselves after a new patch is released.

[00:01:32]And when one of them finishes the actual work on the code, you're back to a finished commit, push, and draft merge request waiting for review.

[00:01:39]Restarting the agent now opens a list of previous sessions, including those you deleted. Even MCP calls that last longer than 2 minutes automatically go to the background instead of blocking you.

[00:01:51]Greater autonomy requires better brakes. Automatic mode now works on Bedrock, Vertex, and Foundry without the need for a connection. If it deviates from course, Claude's automatic reset clears everything after user confirmation. The old default permissions mode is now honestly called " manual". Claude mode was observed to automatically run commands that modify files. This vulnerability has been closed, and destructive commands with undefined variables now stop and ask for permission before deleting.

[00:02:20]Security has undergone the biggest changes this cycle, in a good way. PowerShell 5.1 had a bypass for permissions checking.

[00:02:29]Fixed: Permission rules for individual directories silently granted access anywhere in the tree. Fixed: Invisible characters that masked malicious commands are now neutralized in permission previews. The agent tool is protected against indirect prompt injection.

[00:02:44]Sandboxes can now block access to files containing credentials.

[00:02:47]Docker commands that redirect the daemon now prompt for permission. And the new tool allows Claude to independently end conversations with abusive users and hacking attempts. And now an update for daily work. The fork command copies the conversation to a background session while the original continues, and the subtask replaces the old insession helper. You can combine up to five skills into one line. The rewind command now continues to work, so yesterday's work won't disappear. A new database skill creates graphs and dashboards on demand. Bash mode now has autocompletion of file paths as you type. Side questions now have their own panel. One team returns the last of them. And for Vim users: two-key escape sequences like JJ can finally be assigned.

[00:03:35]The terminal has become faster. Long tables and huge paragraphs used to freeze while rendering— fixed.

[00:03:41]Scrolling no longer jumps in the middle of the stream. Flicker and split terminals are a thing of the past thanks to synchronized output. Markdown tables no longer stop the entire session. Input remains responsive when updating task lists, and threading updates reduces CPU usage by 37% in long sessions. Which brings us to the promise at the beginning: the end of constant surveillance. Here's how it works. Network outages no longer interrupt progress. Temporary errors are repeated with a delay.

[00:04:11]Subagents restricted by request limits no longer disappear without a trace. They report and continue their work.

[00:04:17]Background sessions can withstand stopping, restarting, and even replacing the daemon while it is running. Interrupted agents resume their work independently. So the nightly assembly line that used to require your presence at 2am now only requires a code check over coffee. The move from monitoring agents to checking the results of their work is the real upgrade in this version. One team gives you all of this. Update Claude. So, honest question.

[00:04:44]Agents that commit, push, and independently open pull requests. Is this a dream or a danger you would never allow into your code?

[00:04:53]Choose a side in the comments.