Ive been using CC to automate desktop work for a while now, and it’s been amazing, I wrote an article about it. Smart move from Anthropic to position this as a more accessible tool for non-technical users, I believe there is so much automation value you can get from CC.
Love that you’ve already pushed CC into real desktop workflows! And yes to the risks; if we're giving CC file access, we should be able to understand what they are. Thank you for reading @Ashwin Francis! 🤗
Karo, this is absolutely amazing. I know you mention it in the article, but the thing that would be holding me back from using this more widely with my own work is definitely the risk of prompt injection. I also wonder to what extent Anthropic are going to be learning from all of the ways in which we start using Cowork to then start building new tools and potentially having new advertisement streams as well.
Totally with you: until we have opinionated guardrails and better defaults, “just trust the agent on your real data” is a non‑starter for many people. And teams! Thank you for reading Sam 🤗
Thank you for writing this, Karo. I’m grateful you’re on top of these shifts and translating them in real time.
The speed from “niche coding tool” to “general‑purpose digital coworker” in ten days is wild. What really amazed me me is that this wasn’t a whiteboard fantasy so much as Anthropic used user behavior and had an agent ship its own non‑technical sibling.
This 10-day ship cycle is exhilarating, but it also highlights the emerging 'Authority Gap.'
When AI builds AI at this speed, the friction of development disappears, but the friction of trust increases. As Karo rightly points out, the bottleneck is moving from intelligence to trust.
In my work de-risking B2B content, I'm seeing that the faster we 'Cowork' with Claude, the more we need a human 'Editor-in-Chief' to ensure the output doesn't lose its soul or its strategic weight.
Speed is a tool; 'Quiet Dignity' is the strategy. Great deep dive, Karo.
Really great framing of the Authority Gap. This is it: speed is now almost the easy part, and the hard part is earning the kind of trust where you’d actually let the thing touch your real real work. Thank you for reading Osman! 🤗
The insight about watching behavior over surveys is gold. I noticed the same thing building my personal AI agent (Wiz) - I set out to automate tasks but ended up with something that handles email processing, job searching, blog writing, and relationship tracking. The tool found its own use cases.
Anthropic's framing shift from "Claude Code" (coding tool) to "Cowork" (agency + automation) mirrors what I'm seeing in e-commerce too. AI agents won't just help shoppers - they'll shop FOR them.
The trust question you raise is critical. I had to build verification loops, human approval gates, and memory systems before I felt comfortable letting an agent operate on my files autonomously. That's the real product work.
Yes, the pace of development is changing! I just built a production-ready app in an afternoon so I can try out Vercel…It’s entirely credible that Anthropic built Cowork in 10 days. 😳
> “Code” screams “tech expert only” to anyone who doesn’t see themselves as one.
You know if I could wave a magic wand, I would make it to help people change their self-image in a snap.
Maybe is cause I've been labeled the "tech guy", my family and close friends have never tried doing stuff that is within reach to them cause they don't see themselves as "techie".
Thank God that with AI this is now a myth that's getting busted. 💪
Thank you for sharing, it's a very relatable perspective for many of us. And you're right, ''techie'' might start meaning something new very soon (or even stop being needed as a word at all?)
This might be my favourite question Jose! That's exactly why I push for building with attitude. The long‑term advantage will shift from being “the only one who can build” to being the one who can repeatedly ship MEANINGFUL & TRUSTED things. When everyone gets access to the same hammers, it's the craft that matters.
I agree with the core claim. The bottleneck has moved from intelligence to trust, and that is exactly why we need governance. Trust is not adoption friction. Trust is accountability architecture. The moment a system can execute work, touch files, and operate across time, the question is no longer how impressive it is. The question is who owns the decision when it goes wrong, what the audit trail looks like from intent to action, and how you revoke or roll back when the environment lies to it or it misunderstands you.
Anthropic deserves credit for being unusually transparent about risks like prompt injection and for building containment like sandboxing. That is serious infrastructure. But containment is not governance. Sandboxing can limit blast radius. It does not answer who is liable, what threshold of failure is considered acceptable, or who signed off on that threshold. Ten days to ship is an acceleration story. It is also a verification story, because whatever you do not explicitly test and bound at that velocity becomes debt that accrues in public.
This is where governance stops being a bolt-on and becomes the product. You cannot just tell users “trust us” while you hand an agent write access to their world. You need a system that makes trust legible. Provenance. Receipts. Escalation rules. Human-in-loop for irreversible actions. A clear revocation path. You need bright lines that prevent authority from drifting into places where no consequence can land.
DAS-1 is one example of the kind of line that has to exist if we want this to be safe. Delegated authority without consequence and liability is the failure mode that will show up by default if we do not design against it. If “the agent did it” becomes a normal sentence in an incident review, governance already failed. The advantage is not just speed. The advantage will belong to the teams that can move at this speed while keeping authority, accountability, and liability welded to named humans with receipts.
Completely aligned. I agree with you that sandboxing is necessary but nowhere near sufficient. ''Trust is accountability architecture” is exactly it. Very well said. Thank you for reading Paul!
Exactly, I live for this stuff: where you don’t invent a fantasy use case, but follow the real workflows and then get out of the way. Thank you for reading Michael!
I am completely awestruck how accessible coding is becoming thanks to AI. I honestly never in my life would have thought I'd be building something one day, and now it's looking inevitable. My curiosity is calling me to experiment,
Yes!! This is my favorite part: the people who never self‑identified as “technical” suddenly realized they CAN sketch an idea and watch it turn into something real.
Ive been using CC to automate desktop work for a while now, and it’s been amazing, I wrote an article about it. Smart move from Anthropic to position this as a more accessible tool for non-technical users, I believe there is so much automation value you can get from CC.
I really appreciate you calling out the risks.
Great post as always Karo!
Love that you’ve already pushed CC into real desktop workflows! And yes to the risks; if we're giving CC file access, we should be able to understand what they are. Thank you for reading @Ashwin Francis! 🤗
Karo, this is absolutely amazing. I know you mention it in the article, but the thing that would be holding me back from using this more widely with my own work is definitely the risk of prompt injection. I also wonder to what extent Anthropic are going to be learning from all of the ways in which we start using Cowork to then start building new tools and potentially having new advertisement streams as well.
Totally with you: until we have opinionated guardrails and better defaults, “just trust the agent on your real data” is a non‑starter for many people. And teams! Thank you for reading Sam 🤗
Thank you for writing this, Karo. I’m grateful you’re on top of these shifts and translating them in real time.
The speed from “niche coding tool” to “general‑purpose digital coworker” in ten days is wild. What really amazed me me is that this wasn’t a whiteboard fantasy so much as Anthropic used user behavior and had an agent ship its own non‑technical sibling.
Yes, it's the wildest thing I've heard of recently tbh. It sets the scene for everyone else. Thank you for reading Anna!
This 10-day ship cycle is exhilarating, but it also highlights the emerging 'Authority Gap.'
When AI builds AI at this speed, the friction of development disappears, but the friction of trust increases. As Karo rightly points out, the bottleneck is moving from intelligence to trust.
In my work de-risking B2B content, I'm seeing that the faster we 'Cowork' with Claude, the more we need a human 'Editor-in-Chief' to ensure the output doesn't lose its soul or its strategic weight.
Speed is a tool; 'Quiet Dignity' is the strategy. Great deep dive, Karo.
Really great framing of the Authority Gap. This is it: speed is now almost the easy part, and the hard part is earning the kind of trust where you’d actually let the thing touch your real real work. Thank you for reading Osman! 🤗
The insight about watching behavior over surveys is gold. I noticed the same thing building my personal AI agent (Wiz) - I set out to automate tasks but ended up with something that handles email processing, job searching, blog writing, and relationship tracking. The tool found its own use cases.
Anthropic's framing shift from "Claude Code" (coding tool) to "Cowork" (agency + automation) mirrors what I'm seeing in e-commerce too. AI agents won't just help shoppers - they'll shop FOR them.
The trust question you raise is critical. I had to build verification loops, human approval gates, and memory systems before I felt comfortable letting an agent operate on my files autonomously. That's the real product work.
Related reading if interested: https://thoughts.jock.pl/p/agentic-commerce-ai-shopping-for-you-2026
Can cowork listen to audio recordings on its own or do you need anything additional?
I'm not sure.
I love Claude Code and definitely should try out Cowork as well. Currently only limited to Max users, right?
I think Pro users just got access too. Thank you for reading Wilbert 🤗
A new world of eating your own dog food!
Yes.
Yes, the pace of development is changing! I just built a production-ready app in an afternoon so I can try out Vercel…It’s entirely credible that Anthropic built Cowork in 10 days. 😳
> “Code” screams “tech expert only” to anyone who doesn’t see themselves as one.
You know if I could wave a magic wand, I would make it to help people change their self-image in a snap.
Maybe is cause I've been labeled the "tech guy", my family and close friends have never tried doing stuff that is within reach to them cause they don't see themselves as "techie".
Thank God that with AI this is now a myth that's getting busted. 💪
Thank you for sharing, it's a very relatable perspective for many of us. And you're right, ''techie'' might start meaning something new very soon (or even stop being needed as a word at all?)
Yeah, that'd be pretty cool. It might happen with the remote work vs onsite work thing. In the pandemic remote work became just "work".
Great analogy!
Karo, I wonder how these developments affect people who are building their skills and identities as AI builders.
I was talking with a friend today about a potential issue:
if everyone becomes a builder, will there still be a long-term competitive advantage for those building today?
I’d love to hear your thoughts on this.
This might be my favourite question Jose! That's exactly why I push for building with attitude. The long‑term advantage will shift from being “the only one who can build” to being the one who can repeatedly ship MEANINGFUL & TRUSTED things. When everyone gets access to the same hammers, it's the craft that matters.
That is a very useful answer.
Every builder in Substack should read it.
Thanks
I agree with the core claim. The bottleneck has moved from intelligence to trust, and that is exactly why we need governance. Trust is not adoption friction. Trust is accountability architecture. The moment a system can execute work, touch files, and operate across time, the question is no longer how impressive it is. The question is who owns the decision when it goes wrong, what the audit trail looks like from intent to action, and how you revoke or roll back when the environment lies to it or it misunderstands you.
Anthropic deserves credit for being unusually transparent about risks like prompt injection and for building containment like sandboxing. That is serious infrastructure. But containment is not governance. Sandboxing can limit blast radius. It does not answer who is liable, what threshold of failure is considered acceptable, or who signed off on that threshold. Ten days to ship is an acceleration story. It is also a verification story, because whatever you do not explicitly test and bound at that velocity becomes debt that accrues in public.
This is where governance stops being a bolt-on and becomes the product. You cannot just tell users “trust us” while you hand an agent write access to their world. You need a system that makes trust legible. Provenance. Receipts. Escalation rules. Human-in-loop for irreversible actions. A clear revocation path. You need bright lines that prevent authority from drifting into places where no consequence can land.
DAS-1 is one example of the kind of line that has to exist if we want this to be safe. Delegated authority without consequence and liability is the failure mode that will show up by default if we do not design against it. If “the agent did it” becomes a normal sentence in an incident review, governance already failed. The advantage is not just speed. The advantage will belong to the teams that can move at this speed while keeping authority, accountability, and liability welded to named humans with receipts.
https://github.com/forgedculture/das-1
Completely aligned. I agree with you that sandboxing is necessary but nowhere near sufficient. ''Trust is accountability architecture” is exactly it. Very well said. Thank you for reading Paul!
Anthropic didn’t just build a product, they observed real user behavior and let it guide the design
Exactly, I live for this stuff: where you don’t invent a fantasy use case, but follow the real workflows and then get out of the way. Thank you for reading Michael!
I think anthropic change the game with this one
Right? The potential is huge.
I know i wish i didnt have to wait for it
And thank you for reading Alexander! 🤗
I am completely awestruck how accessible coding is becoming thanks to AI. I honestly never in my life would have thought I'd be building something one day, and now it's looking inevitable. My curiosity is calling me to experiment,
Yes!! This is my favorite part: the people who never self‑identified as “technical” suddenly realized they CAN sketch an idea and watch it turn into something real.
Thank you for reading AI Girlboss!!!
Anthropic didn’t just build a product, they watched user behavior, removed barriers, and let AI execute autonomously
Completely agree. This is the pattern I want to see more of: observe behavior → plan how to remove barriers → let AI run.