Claude Code Auto Mode Explained: Safer Than Skipping Permissions

By Beau Johnson··8 min read

Quick answer

Claude Code auto mode uses an AI safety layer to check every action. Safe ones, like writing files or running tests, go through automatically, and risky ones get blocked. It's the middle ground between clicking Allow on everything and running dangerously-skip-permissions.

In this video I explain auto mode for Claude Code. It's Anthropic's answer to the permission problem, where you either click Allow over and over or hand the AI full control with dangerously-skip-permissions. I recorded this in March 2026, when auto mode had only been out for about a day as a research preview. Tools change fast, so check Anthropic's current docs before you rely on any detail here. This one is for anyone running Claude Code or building AI agents who's tired of babysitting every action.

What I cover in this video

  • 0:46 The permission problem: Claude Code stops to ask before writing files, running bash commands or installing packages
  • 1:31 What dangerously-skip-permissions does and why it lives up to its name
  • 2:18 How auto mode works as the middle ground
  • 4:34 My setup on the Mac mini and how auto mode changes it
  • 5:20 What Anthropic hasn't explained yet, and the TechCrunch callout
  • 6:52 The bigger picture: a step toward truly autonomous AI agents
  • 8:23 Three things builders should do right now
  • 9:08 Anthropic's 48-hour shipping spree and the security angle

The problem: babysit it or let it run wild

Normally Claude Code stops and asks you for permission every time it wants to write a file, run a bash command or install a package. Allow, deny, allow, deny. If you've built anything serious with it, you know the feeling of walking away for 10 minutes, coming back, and finding it sitting there waiting for you to click Allow. It did nothing the whole time.

The way around that has been dangerously-skip-permissions. Flip it on and Claude Code does whatever it thinks it needs to do: edits files, runs commands, installs things, deletes things, all without asking. It is much faster. But you're giving an AI full control of your terminal. If a prompt injection attack is hidden in some code it reads, or it just misreads what you wanted, it can do real damage. I've heard stories of people coming back to find their project in a completely different state or even a different language, with files deleted, dependencies changed or config files overwritten.

So for months the choice has been to babysit every action and lose hours clicking buttons, or let it run wild and hope nothing breaks. Nothing sat in between.

How auto mode works

With auto mode on, Claude Code still checks every action before it runs. The difference is that it doesn't ask you. An AI safety layer decides whether each action is safe or risky, in real time.

  • Safe actions just happen. That covers things like writing a new file, running a test or installing a normal npm package.
  • Risky actions get blocked. That covers things like deleting files, running unfamiliar scripts or anything that looks like a prompt injection attack. They don't get flagged for your approval. They're blocked outright.

The way I think about it, it's like having a smart assistant who knows the difference between "yeah, go ahead and install the testing library" and "wait, why are you trying to delete the production database?"

When I recorded this, auto mode was a research preview for Team and Enterprise users, with API access expected within a few days. It only worked with Sonnet 4.6 and Opus 4.6, and Anthropic recommended testing it in a sandbox environment.

What Anthropic hasn't told us

Anthropic hasn't said exactly what criteria the safety layer uses to decide what's safe and what's risky. TechCrunch called this out, saying developers will want to understand it better before adopting it widely. I agree. If an AI is making decisions about your codebase without asking you, you want to know the rules it follows. Still, the careful rollout (research preview, limited to Team and Enterprise at first, sandbox recommended) makes me trust it a bit more, not less. Anthropic did the same with Claude Code review, Dispatch and computer use: ship a preview, get real feedback, then improve it.

Why this matters for AI agent builders

If you run automated coding workflows or build with OpenClaw, permissions have been a real bottleneck. You set an agent to build something overnight and wake up to find it finished maybe 30% of the work because it kept hitting permission walls. Your other option was to skip permissions and gamble.

In my own setup, I run Claude Code tasks through my agent on a Mac mini. Solomon writes the code for my mission control dashboard and runs on Opus 4.6. Creating files, running tasks and updating the database schema all needed approval. So I either babysat it, which defeats the point of automation, or skipped permissions, which made me nervous. Auto mode is meant to let the routine stuff happen and catch the scary stuff.

The bigger picture is that AI agents can think, reason and write code, but they're still on a leash with a human approving every action. That's like hiring a smart employee and making them ask permission before opening every door in the building. It works, but it doesn't scale. If the safety layer works as well as Anthropic says, agents could work for hours or even days on complex projects without you stepping in for routine tasks. I still check my agents' work, though, and I think you should too.

This also fits a pattern. Within about 48 hours, Anthropic launched computer use, Dispatch for Cowork and auto mode. All three point toward more autonomy. Computer use lets Claude control your desktop, Dispatch lets you assign tasks to AI agents from anywhere, and auto mode lets Claude Code work without constant babysitting. Anthropic is building the full stack of autonomous AI tooling while trying to balance it with safety: speed when it's safe, control when it's not.

Three things to do right now

  1. Turn it on, but in a sandbox first. If you have a Claude Team or Enterprise subscription, try auto mode on a test project, not your production code. Give it a few tasks and watch what it allows and what it blocks, so you get a feel for where the safety line is.
  2. Rethink your workflow. If you designed your automations around the permission bottleneck, you may be able to simplify them. Instead of building elaborate approval chains, let Claude Code handle the routine work and only send you the truly critical decisions.
  3. Watch what Anthropic does next. This is a research preview, and the rules for safe versus risky will get refined based on user feedback. If you adopt it early and report issues, you're helping shape how it works for everyone.

Key takeaways

  • Auto mode sits between approving every action by hand and running dangerously-skip-permissions.
  • An AI safety layer checks each action in real time. Safe actions run automatically and risky ones are blocked, not sent to you for approval.
  • At launch it was a research preview for Team and Enterprise users, worked only with Sonnet 4.6 and Opus 4.6, and Anthropic recommended sandbox testing.
  • Anthropic hasn't shared the exact safety criteria, so test it on non-production projects first.
  • Auto mode is a step toward agents that can work for long stretches on their own, but I still review my agents' work.

FAQ

What is Claude Code auto mode?

It's a Claude Code mode that uses an AI safety layer to check every action before it runs. Safe actions, like writing a new file or running a test, happen automatically. Risky ones, like deleting files or running unfamiliar scripts, are blocked.

How is auto mode different from dangerously-skip-permissions?

Dangerously-skip-permissions lets Claude Code do anything without asking, including deleting files or overwriting configs. Auto mode still checks each action and blocks the risky ones, so you get the speed without giving up all control.

Who can use Claude Code auto mode?

When I recorded this in March 2026, it was a research preview for Claude Team and Enterprise users, with API access expected within a few days. It only worked with Sonnet 4.6 and Opus 4.6. Check Anthropic's docs for current availability.

Is Claude Code auto mode safe to use on production code?

Anthropic recommends testing it in a sandbox environment, and it hasn't shared the exact rules the safety layer uses. I'd start with a test project, watch what it allows and blocks, and only then use it on real work.

Chapters

  1. 0:00The permission problem with Claude Code
  2. 2:00What dangerously-skip-permissions actually does
  3. 3:30Auto mode explained
  4. 5:15Why this matters for AI agent builders
  5. 7:00My setup and how this changes my workflow
  6. 8:30What Anthropic isn't telling us (TechCrunch callout)
  7. 10:00The bigger picture: autonomous AI agents
  8. 12:00Three things builders should do right now
  9. 13:30Anthropic's 48-hour shipping spree
  10. 15:00Security angle and honest take

Transcript

Read the full transcript

0:00Anthropic just dropped another update that I believe is going to change how builders build forever. They just released something that can use with Claude code and it's called auto mode. And if you've ever used Claude code and you're sitting there clicking around and clicking allow over and over again and Claude code is trying to do its job, this is the update you've been waiting for. They've been shipping a lot of stuff. It's been hard to keep up, but I wanted to make a dedicated video for this because it's going to help speed up your Claude code workflow. So, they just launched it yesterday as a research preview. Every major tech outlet is covering it. TechCrunch, 9to5Mac. This is a big deal. I'm going to explain what's happening here because I think a lot of people are going to misunderstand this. So, right now when you use Claude code, every time it wants to do

0:46something like write a file, run a bash command, or install some package, it stops and it asks you for permission. Allow, deny, allow, deny over and over again. It can be pretty annoying. And if you've ever tried to build something that's serious with Claude code, you know exactly what I'm talking about. You walk away for 10 minutes thinking it's working, doing its job, come back, and it's just sitting there waiting for you to click allow, and it didn't do anything. So, you just lost 10 minutes. Now, there has been an option called dangerously skip permissions, and while you could do that certainly, you know, it's exactly how it sounds. You know, you flip a switch and Claude code just does whatever it wants and no questions asked it, right? Files, run commands, install things, delete things, whatever it thinks it needs to do. And yeah, it's

1:31faster, and that's been the out-of-the-box way to just solve this problem. It is way faster, but it's called dangerously skip permissions for a reason. You You just really don't know what it could be doing. It's actually dangerous. You know, you're giving an AI full control of your terminal, and if something goes wrong, if there's a prompt injection attack hidden in some code that it's reading, or let's just say it misinterprets something, it can do real damage. And I've heard stories of people who have turned this on, came back to their computer and find the project in a completely different state or in a completely different language. Files deleted, dependencies changed, or config files overwritten. So, for months now, the choice has been this, you know, either you babysit every single action and lose hours clicking buttons, or you let it run wild and hope nothing breaks.

2:18But, there has been no middle ground until now, up until yesterday. Auto mode is that middle ground. Here's how it works. You know, when you turn on auto mode in Claude code, it still evaluates every action before it runs, but instead of you asking it or it asking you, it uses AI a safety layer basically to decide whether the action is safe or risky. So, safe actions like writing a new file or, you know, running a test, installing a normal NPM package, those just happen automatically. You don't have to do anything. Risky actions like deleting files or, you know, running unfamiliar scripts, anything that might like, you know, might like be a prompt injection attack, it's going to be more hesitant to do that. So, those get

3:03blocked immediately. They don't just get flagged for your approval, they just get blocked. So, you're not really choosing between speed and safety anymore. You get both. The AI is making judgment calls about what is safe and what is not, and it's in real time for every single action. Now, let me tell you why I think this matters so much, especially for builders like you and me, and this goes beyond just convenience. You know, if you're running AI agents, if you're building with open claw, if you're doing anything that's automated that requires coding workflow, the permission bottleneck has been a real problem. You know, you set up an agent to go build something overnight, right? You wake up, you check on it, it built maybe 30% of what you expected because it kept hitting those permission walls. So, either that or you skip permissions entirely, and you're basically gambling

3:49at that point cuz you don't know what you could be coming back to. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full open claw course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or open claw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have

4:34over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. So, if you think about my setup for a second, you know, I run Claude code tasks through my agent on the Mac mini. There is a generate scripts, Solomon writes the actual code for mission control. He is running on Opus 4.6. So, if Solomon needs to run and create files or run tasks and update the database schema, every one of those actions requires approval. That either means I'm sitting there babysitting the agent, which defeats the whole purpose of automation, or I'm running with permission skip, which makes me more nervous if I am going to run and go down

5:20that road. Auto mode fixes this. The routine stuff just happens. The scary stuff gets caught. It's like having a smart assistant who knows the difference between, "Yeah, go ahead and install the testing library." versus, "Wait, why are you trying to delete the production database?" That's the kind of judgment call that does save you from a disaster, especially if you're just getting into this AI space. Now, I want to be clear about this. Anthropic hasn't told us exactly what criteria the safety layer uses to decide what's safe and what's risky. TechCrunch actually called them out on this in their article. They said developers are going to want to understand this better before adopting it widely. 100% agree. When you're giving an AI the power to make decisions about your code base without asking you, you want to know what rules it's following. And here's my take on that.

6:06The fact that Anthropic is being cautious about this, releasing it as a research preview first, limiting it to team and enterprise users initially, and then recommending you use it in a sandbox environment, that actually makes me want to trust it a bit more, not less. They're not just throwing this out and saying, "Good luck." They're saying, "Hey, test this carefully, give us feedback. Let's make sure this works before it rolls out to everyone." Could you imagine the implications if they did that and it didn't do what it intended to do? You know, and that's Anthropic's playbook. They're rolling it out slowly. They did the same thing with Claude code review, same thing with dispatch, and now the same thing with computer use. So, ship it as a preview first, get the real feedback, and iterate. By the time this hits the general availability, it's

6:52been battle-tested by thousands of developers. I'm going to talk to you about the bigger picture because I think this is what most people are going to miss right here. Auto mode is not just a convenience feature. It's the step towards truly autonomous AI agents. And if you think about it, when open claw releases an update, Anthropic is already shipping one step ahead. So, there is a huge battle going on in this space, if you hadn't noticed. You know, right now AI agents are powerful, but they are leashed. And a lot of the times when you set them up, you know, you have them run in a certain way where you're still checking the work. And that's how I still do mine. You know, I still check their work. And I think you still should, too. You know, they can think, you know, and they can reason, and they can write code, but they need a human in a loop to approve every action. We're

7:38still at that stage. That's like hiring a really smart employee and then making them ask permission before they open every door in the building. It works, but it doesn't scale quite yet. Auto mode starts to solve that. It, you know, gives the AI a sense of judgment about what actions are appropriate. And if Anthropic can get this right, if the safety layer, you know, actually works as well as they claim, then we're looking at a future where AI agents can actually work for hours and maybe days on complex projects without needing human inter- intervention for routine tasks. We're getting there. You just point them at a problem, walk away, and come back to the results. I I feel like we're seeing a massive shift in this whole agentic AI and agents as a service workspace. You know, we're going from AI

8:23as a tool that you use to AI as a teammate who uses tools. And the permission problem was one of those biggest barriers to getting there. Now, for the builders watching this, and here's what I would actually do with this. First, if you have a Claude team or enterprise subscription, turn on auto mode. Get going. But, use it in the sandbox first. Don't run it against your production code right away. Create a test project, give it a few tasks, watch how it handles those permissions, and see what it allows through and what it blocks. Get a feel for, you know, the safety boundary and what Anthropic has put in place. The second thing is you need to think about how this changes your workflow. You know, if you've been designing your AI automations around the permission bottleneck, those designs might need to be updated. You might be able to simplify things. Instead of building elaborate approval chains, you

9:08might be able to just let Claude code handle the routine stuff and only surface the truly critical decisions to you, you know? And pay attention to what Anthropic does next. It's hard to be keeping up with what they've been shipping every day. You know, this is a research preview. It's going to evolve eventually, and the criteria for what counts as safe and risky are going to get refined based on the user feedback. So, if you're an early adopter and you report issues, you're literally shaping how this feature works for everyone else. I'm going to also connect this to what I've been seeing from Anthropic lately because the timing is not a coincidence. In the last 48 hours, Anthropic launched computer use. They launched dispatch for co-work, and now auto mode. Three major features in 2 days. They're on a huge shipping moment

9:54right now, and all three of these features point in the same direction. More autonomy for AI, computer use lets you use Claude to control your whole desktop, dispatch lets you assign tasks to AI agents from anywhere, auto mode lets Claude code work without constant babysitting. Does this sound familiar if you're an open Claude user? They're building the full stack of autonomous AI tooling now. And if you're paying any attention, you can see where this is going. Within the next year, I predict sooner, we're going to have AI agents that can handle entire workflows end to end. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. I built something called content machine. It's 10 AI agents that run on the open Claude

10:39orchestration, and they handle everything. Scripts, thumbnails, X posts, [music] blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up, and [music] the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe [music]

11:24and join the community. So, you plug in your own thing, and it molds it to you. It learns how you talk, and it writes, [music] so it doesn't sound like AI slop. $97 one time, it's not a subscription. I'll put the link down below, and you'll thank me later. So, code review, deployment monitoring, bug fixing, all without a human in the loop for every step. This is why I keep saying if you're not building with AI right now, you're going to be behind. This isn't just a fun demo anymore. This isn't speculation. These features are shipping. They're available today. The tools exist right now for you to start building things that would have been impossible a year ago. I was talking to someone in our community last week who said they had been on the fence about learning Claude code. They thought it was just another hype cycle. And I told them, "Th- This is more looking at like

12:11the shipping velocity. Look at what Anthropic's doing. And look how open Claude and Anthropic and all of these companies like Google are trying to get as many features to serve their audience as fast as possible." You know, look at the things that are coming out every week. It's hard to keep up. This isn't slowing down. It's actually in accelerating in its exponential phase right now. And the people who start building now are the ones who are going to have the skills, the workflows, and the muscle memory when it comes to this stuff really taking off. You don't want to be learning the basics when everyone else is already running autonomous AI agent teams. That's the thing about timing. You can't learn experience. You can't compress the trial and error of building something real. My favorite thing that I've picked up in this last couple years is learning how to use

12:56Claude code. Breaking things, learning what things did and why the agent did it. That's my biggest takeaway is that if you're don't know anything about Claude code, dive in and start building something. And really, that's the only way to get good at it is just through repetition. You know, to start building, to start making mistakes, start figuring out what's working for your specific situation. And the tools have literally never been this good. Now, I also think there is a security angle here that's worth talking about, especially, you know, with what we've covered earlier about the Claude hub malware situation, and auto mode is Anthropic's way of saying like, you know, we need you to know the speed here. And, you know, we're not going to sacrifice safety to give it to you quickly. You know, and that matters. And a lot of other tools in this space are just taking the

13:41opposite approach. They just let the AI do whatever, move fast. And we've seen what happens when you move fast without guardrails. You get a malicious skill on Claude hub. You get a supply chain attack. You get people losing control of their codebases. Anthropic is trying to find a balance here. Speed when it's safe, control when it's not. And that is the best approach that I respect the most from Anthropic. Because this stuff is powerful enough to do actual real damage if you're not careful. So, yeah, like Claude code auto mode research preview, it's available now for team users, enterprise, and the API is coming in the next few days. Only works with Sonnet 4/6 and Opus 4/6. Anthropic recommends sandbox environments for testing, and I think it's one of the most important features they've shipped in the last week. So, if you want to learn how to actually get your first AI

14:26agent going, and how to set up Claude code properly, how to actually ship products with this stuff, come join us in the community. 142 members on day 14 since we launched the community. We do six live boot camp calls every single week to help you take the next step in your AI journey. I'll put the link down below. And if you guys aren't subscribed to the channel already, we release four to six videos every single day to help you keep up with the AI news. Hope you have a blessed day, and we'll see you in about an hour or two.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading