Claude Tag Looks Useful. But It's Dangerous

So Claude launched a feature that looks incredibly useful. It's called Claude Tags, and you can basically have your Claude talk to your apps inside of Slack, which means you can bring it into a conversation, you can have it interact with teammates, and overall, you should be more productive with it. But this is why I think most people in business need to be paying attention to this. This might be the start of AI companies routing, remembering, and executing your work.

Okay, so this is Claude Tag. You can see it on my screen. and basically it says, okay, it's a new way for teams to work inside of Cloud. So Cloud joins as a team member and just access the channels.

And for example, you can connect different things such as your HubSpot, for example, like your CRM or your project management tools. And you can see here, this person's interacting. This is in a product engineering channel where they're interacting with Cloud. And then it's basically, it can do data pools for you.

It can even schedule cron jobs and it can read and write for you. So it can update Asana. It can update linear for you, or we can pull data from HubSpot. you can pull data from gong right and so you know they say they've been working like this for a while inside of inside of slack and so even over here you schedule exports line in this afternoon okay you know launch date doesn't move it's now in beta but the blog post and the beta invite email didn't mention do you want me to add both okay great great it's added over here that's the change of social threat so basically what this is is this isn't necessarily new to those of you that might have been using open claw or hermes so this is claude directly moving into the the workspace right And you can see like, there's obviously a lot of interest here.

You have 15.4 million views on it. What I would say is you might be asking yourself, okay, well, how is this different than using Claude? Well, before Claude was more, it was more single player, right?

Meaning that you're opening up your own instance of Claude, your own instance of Claude co-work, Claude code, whatever it is exactly. You are, you're executing on your own skills. Maybe sometimes you're sharing your skills with other people, but when you are able to sit in a channel and you're able to talk through an idea and then someone else has maybe seen your idea, maybe they come up with another question and then Claude detects that. Well, that's good because what ends up happening is that you're able to move a lot faster because you're communicating a lot faster and the exchange of ideas is flowing through it, which means that if that's happening, other people are learning from it.

It's kind of spreading virally throughout your organization. That ultimately means you can get things done a lot faster. And that's certainly what we've seen using our single brain internally at my company. And so I think it's really good that they're doing this.

And so it's cool that it can sit inside of Slack. It's cool that it can it can follow team conversations. It's cool that it can remember things. It's cool that it can access tools.

It's cool that it can act asynchronously. And I would just say one thing is most people haven't considered the trade-offs to using something like this. So I think it's great. It's going to get more people working in this fashion.

And my prediction has always been that most companies are going to work in this way in the next 12 months or so. I think that timeline has actually accelerated now, now that Cloud has released this. And I'm pretty sure OpenAI will have something like this coming out soon. So the question you have to ask yourself again is, Like, do you start using this right now?

And then what are the trade-offs? But then also, what are the potential solutions here if you don't want to use something like this, right? And that's something that we'll talk about if you stay till the end. So something to keep in mind is the way you were using Cloud before, single player is more as a, Cloud is your model.

It's a frontier model, right? You just want to think about this. The moment that it becomes a shared coworker, it becomes a different relationship. It's no longer just a model provider.

It's more so it's an operating layer in your company, which is why we call ours single brain. It is very much that company brain that people are talking about because you want this memory that compounds with you over time. And one thing to call out, too, is that the way Claude Tag works is that it isn't charging you on a proceed basis. It's actually charging you on a token basis.

So the more you use it over time, and I even saw the setting yesterday, it's like, oh, do you want it to just be unlimited? No, I don't want it to be unlimited, right? But it like do you want it to be a month a month Like what limit do you want to set to cap on this thing And you can set it for Opus 4 or Opus 4 which is interesting to me because they don have Opus like Sonnet for example You can use one of the cheaper models to use it right And so they can kind of dictate whatever they want there.

And so just keep in mind, if you're thinking about switching over to this, that's one of the risks that we'll talk about in a second, right? So, all right, so let's talk about some of the risks here. And then some smart people here are calling on a couple of things, right? So the benefit here, like I mentioned, is, okay, Claude is multiplayer.

It learns over time, it takes initiative, and it works asynchronously. You know, what Arvind is saying here is that it seems useful, but it's very dangerous. It's a dangerous bargain for enterprises because of the pricing model and the risk of locking. So the four big changes together mean that you interact with Claude as a co-worker instead of a tool.

And what happens is the same Claude instance for everyone instead of each worker soaks up tacit knowledge without you telling it, acts on its own, and does so asynchronously. So it's all clearly very useful, but it completely, it flips the interaction diagram, right? So pricing, we kind of talked about a little bit. It's not like it's pricing you on a per seat model.

It's pricing you on usage so it can compound the more you use it. Also, it talks about lock-in, right? So Cloud maintains its own memories and this new way of working. The human team members can't see and edit them.

So system administrators presumably can, but they have other things to do. So tacit knowledge thus goes from a weakness of AI agents to a major strength. It seems inevitable that as teams and orgs start to use Cloud this way, it will become the main queriable repository of all their tacit knowledge. That's a single brain piece again, right?

Creating dependence and stickiness. Effectively, Cloud is a coworker that you can't fire without every team losing workflows and know-how. So the moment you take this away, it's not going to be good for your company, right? So you basically are getting locked in, not just from a model standpoint.

I wouldn't really worry about that too much, but more so from a knowledge lock-in standpoint, right? And so then it says, by the way, it seems to introduce a new and pervasive security risk since cloud can be integrated into private channels as well and can be given access to repositories and tools even if the users in that channel don't have access to them. Okay, that can be really dangerous. Imagine you have a bunch of contractors in a channel.

Maybe there are certain things that they should be seeing. Okay, well, now they have access to it, right? They can call on whatever they want. By the way, if you're enjoying this video right now on company brains, we're talking about single brains, we're talking about cloud tag.

You have to check out single brain because single brain is our managed company brains that we make for our clients that live inside Slack and Microsoft Teams, where we're able to help you route tokens to the right model. We are also able to help you compound your memory, your decisions that you make as a company and let that compound with you because that's the most valuable piece. So if you want to check it out, just go to singlebrain.com.

That's with a B, and we'll see you on the other side. So Anthropic has introduced an interesting but complicated access control mode to handle all of this, but I'm not sure I trust people to understand and implement it correctly. So there is a setting in there where it can be only for your full-time employees, but what are the odds that they're thinking about this stuff? They're not thinking about this at all, right?

That's what the reality is. Most people aren't going to be thinking about prompt injection or threat protection at all, right? So I think the world is going to continue to move in this direction. but you know the issue is that you don't want to be just locked into into one vendor right and so ethan molek here also tweeting so decisions about how to use ai in your organization are increasingly organizational design and state and strategy decisions not it choices so how do you integrate agents into your firm that's one thing to think about what intelligence will you outsource and what are the boundaries of the firm and what is the role of people okay so and then this this guy over here ashwin this will be the last piece over here i think these are all good things to consider.

And then I'm going to talk about what your solutions are to moving forward here, right? So Claude Tag is a Trojan horse, not because Anthropic is doing anything evil, and I don't think they're trying to either, but because the incentives are obvious, right? So day one, this looks like a great feature Tag Claude in Slack let it follow the thread remember context connect to do breakdown tasks chase work and act like a teammate right So the moment where AI vendor becomes a shared coworker it stops being just a model provider It starts becoming the place where work is interpreted remembered routed and eventually executed, right?

So context login is what it's called. You're now renting your company back from them. And so all these things over here, right? So agents can be copied, models can be swapped, but the memory of how your company actually works is much harder.

So that's why you want to be having your company brain on your own infrastructure. And so you can go into convenience now. Everyone's like, oh, I'm already using Cloud right now. I might as well do this.

And we love using the Cloud models. Definitely think they're, I mean, we pay for the Teams accounts as well. But I would just say this, if you think about go rent the intelligence from people, but then ultimately, in terms of what you need, your company memory, storage, governance, permissioning, all that stuff, ideally infrastructure that is controlled by you or like another area, right? Maybe it's a managed service by somebody else.

That's how I think about this. Now, in terms of how you want to execute on this, right? Simple. You can set up your own version for your company.

You can set up your own version of NemoClaw. And so you can pull from Hermes there, you can pull from OpenClaw. And that's what we do for our clients. We just use NemoClaw for them.

and then we host our stuff on AWS or we host it on their infrastructure. And that way, it's their brain that compounds. Then that's what we call a single brain. But when you're able to swap out the model, okay, oh, GLM just came out.

The new version of GLM just came out. Okay, oh, wow, these open source models are getting really good. Okay, well, let's have a version of that. What about token routing?

How do we think about that? What about token optimization too? Which is a token routing is a form of token optimization. What about permissioning?

What about governance? All these things. Oh, okay. Well, maybe for the top 10% of tasks, maybe let's go ahead and use the latest version of like a cloud frontier model or open AI frontier model.

That's fine, right? But you want to have more control there. So that memory piece, that is kind of the foundation of what your company is. And then the intelligence you can kind of swap out, right?

And so when I think about the memory of your company and I think about the skills that you have, it's the decisions that you make that compound over time. That's ultimately what helps you compound your business a lot faster. So all that to say, I think this is, I'm not saying Anthropic is trying to be misanthropic here. I don't think they're trying to be evil.

I think they're making a business decision. And I think most people will use Claude tags. But I think if you're watching this right now, if you want to have control of your destiny, it's probably within your own best interest to use something. If you want to have autonomous agents being set up, go use a Nemo Claude and then pick one of the harnesses and then run on NVIDIA's Enterprise runtime.

And then, you know, for us, Sock 2 compliance is important for our customers, right? Or setting up on their infrastructure. But the stuff that I use internally, like, yeah, it is a version of Hermes. And we have memory that compounds over time.

We have all these skills in there. We have a skills dojo for the team. So I actually also wanted to give you a practical checklist on how you can get started with this, okay? Step number one, you want to pick a high-value workflow.

So it could be like ad creative production. It could be around social media, content generation, about AEO, SEO in general. It can be around sales. It can be around weekly reporting.

It can be around just analysis, reporting with your team, right? So these are all workflows that might be high leverage for your business, right? Because I'm talking about it as it relates to revenue because marketing revenue, that's kind of what we focus on more. And step number two is you want to map out the connectors here, right?

And so what do your agents need? Do they need Google Drive? Do they need, obviously, if they're working inside of Slack teams, you're probably going to need that if you want to do something like this. Are you going to use some form of sales intelligence?

So Gong, your CRM. So the more connectors you map and I would say just focus on the critical connectors the more useful your brain is going to be You can call in from Granola too if you want and there an API for that Calling in from Google Drive And keep in mind permission is going to be different at every single company If you're a smaller company, you're going to have the ability to do more with this. Step number three would be defining the memory. So what is it that you want it to remember?

Do you want it to remember your brand voice? Do you want it to remember the decisions that you're making over time. So for example, I log my decisions inside of Obsidian. So you can see it logging over time and it gives more weight to recent decisions versus decisions I might've made three years ago because time changes, right?

Information changes over time. Workflow rules, examples, all that type of stuff. Permissioning, step four is also very important. You got to think about who's able to access what.

You got to think about who can actually approve. You should also think about what data can be exposed. So these are all things that you need to consider. And I I highly recommend that you're working with someone on this.

And again, if you're just starting out and maybe it's yourself, that's totally fine. Number five, if this is relevant, you got to think about model routing. And so the larger the company is, the more that you're spending on tokens. You got to think about token routing, got to think about token optimization.

So maybe 80% or 90% goes to an open source model. 10%, 20% goes to frontier models, right? And then you can choose which ones that you like to use. And then six, you want to think about cost control too, right?

So talk about token budgeting, token optimization. We want to talk about logging. We don't think about approvals. We want to look at usage.

And in step seven, I would say you want to roll it out to maybe very specific teams first, maybe one or two teams. And then once you see that they can't live without it, then you can consider rolling it out to other teams. But don't just try to roll it out to the whole company at once because that can be a mess. I'm just going to give you an example on how I'm using it.

So literally this video, I literally ran through, I have a content machine skill. Okay. So I have a skill over here and it's interviewing me. And then by the way, I use another skill to prepare this for an X long form post.

So I kind of went back and forth with it. And I actually have two videos that I want to make. And it started interviewing me and going back and forth. I actually share some resources in here as well.

And so we also figured out the packaging here, too. So it knows what my preferences are. And so this part, you might be thinking, okay, that's not that special. But it's cool that I, you know, my skills are very portable, right?

Sure, you know, you can talk to your controller. But I also want to look at, you know, what are examples of stupid spend that we have, right? And then so for example, I can pull from your QuickBooks, your build.com, or maybe using something else, like we've actually moved over to ramp because they have an API.

So the more APIs that you have, the more connectors you have, the more you can ask it to do things, right? Okay, as an example, here, I'm putting a dinner for nine to 10-figure AI operators that are at these nine to 10-figure companies, right? And so my outreach over here, I'm like, okay, this is what the outreach looks like. Can you use the Apollo API to figure out who else would be a good fit to invite to this, right?

So I have the Apollo API, I have Clay as well, and you can see these are people over here. So we're going to blur out all these names over here, obviously, but these are all the people to invite, right? And I can kind of sort through the list saying, hey, I like these people here, go find more people like this. And so again, the more connectors you have, the more useful this thing becomes.

You can also say, hey, for example, when we're working with our single brain clients right now, it's like, oh, can you look at the last 30 days or so using Gong and using Granola and also using HubSpot and see what the general sentiment is of these clients and give me examples of what they're saying that proves that they're happy or unhappy. be? Or what are some trends that they're looking for? Like how, how else should we think about working with these people?

Right. And so again, you don't need to wait for that stuff. Anybody that starts to work with us on the single brain side, the moment they get one chat in, it changes everything. So am I saying Claude Tag is evil?

I'm not saying that Claude Tag is evil, right? What I'm saying is this is useful, but you got to consider what the trade-offs are if you're going to sign up for what it is that they have based on what I mentioned. So yeah, that is the video. Hope you enjoyed it and we'll catch you in the next one.