Anthropic launches Claude Tag, letting the AI assistant act as an autonomous coworker in shared workspaces like Slack

"A coworker that remembers everything and bills by the thought." Wish I'd thought of this phrasing myself!

Another thing many people pointed out is that you can easily build this yourself on top of Claude Code (or any other agent) instead of getting locked in. 💯

I'm particularly curious to see how quickly open-source scaffolds will replicate this. Open-weight *models* may be a few months behind the frontier, but fast-following scaffolds can adopt new ideas in a matter of days.

As a big fan of Claude, hate to say I’m generally pessimistic of this direction. Quite literally human distillation at scale. I saw Karpathy in his post mentioning that everyone can tag Claude and contribute. This sadly reminds me of the early era of LLMs where suddenly “everyone can do everything”, yet the reality is that there’s a certain degree of slop now happening even more pervasively at a level that’s even harder to catch or for anyone to easily argue against, and somebody will pay the price for it eventually

@random_walker This was built earlier as one example.

@random_walker There is probably a huge startup opportunity to solve the problem of "getting your employees to use tokens in a way that is net-positive value to your company."

@random_walker coworker is the scariest framing yet

because now its a coworker that remembers everything and bills by the thought

@random_walker Nothing like letting Anthropic spy on and monitor you so it knows if you should be on @DarioAmodei naughty or nice list.

An outage from Anthropic and your entire business is gridlocked.

Yeah, no thank you.

And this is why we’re building Valet AI teammates in Slack:

1) No lock-in → Your agent develops versioned “source code” you can export and take to any other harness. 2) Least-privilege access control by default → You control the domain model and the tools agents have access to. 3) Inference you control → Lower-cost models and on-prem options available. 4) Your agent isn't limited to slack - the same agent can be an MCP or a website - whatever interface your human needs

Try us out at http://valet.dev

@random_walker Agree with all of this. That being said, my hope is that it's only a matter of time before the open-source options follow -- with GLM 5.2, I suspect Anthropic's innovation in this space won't actually accrue benefits to them for very long at all

I couldn't agree more with every single thing that you had there. It is dangerous because without people realizing it, they are building more and more and more dependency on somebody that could at any moment ban you or take that newfound superpower away from you. Not only are they going to take away incredible knowledge company but also your autonomy to operate without constraints.

@random_walker I do not understand why technology teams are accepting anything other than better models from frontier labs. This is just a mechanism to vendor lock, and a huge red flag.

@random_walker Nicely said!

IMO internal agents and intelligence systems are core enough to a company that each company should own as much of it as they can

It's the same decision for whether you hire an employee or outsource some functions to a BPO

This is a very real concern. If your team's knowledge lives inside an AI's private memory, invisible and uneditable, that's not a feature. It's a new kind of lock-in.

Your knowledge should live in YOUR workspace: in your tasks, Docs, and conversations. You should be able to audit everything AI does (and why). You should be able to pick your own model and leave whenever you want.

That's how we built Brain. Not because it's trendy, but because the alternative is exactly what you're describing.

@brhkim Yup.

@random_walker Presumably you’ll be able to download the memory and feed it to another AI company if you want to switch

@random_walker To your point, so much of what Anthropic is doing with Claude products is to continuously obfuscate how the systems work -- which is only *ephemerally* good for UX. When things go wrong, that lack of transparency will require an enormous reckoning

@random_walker as someone with often close to admin-level permissions and who is often thinking about higher level goals, i feel uncomfortable about claude becoming a permanent coworker

i’d much rather use something self-hostable which we can control, with claude being the underlying engine

Agent coworkers sound cool until you realise they often have broad access to sensitive data, and you don’t always have end‑to‑end visibility into where that data is processed or retained. Policies and certifications help, but they don’t guarantee that sensitive data isn’t stored or inspectable by the provider. Combine that with prompt‑injection risk, opaque behaviour, and hard‑to‑control token spend, and these systems can become more liability than asset, especially in highly regulated environments.t, especially in highly regulated verticals

@random_walker The problem is that I understand claude and anthropic and can get my mind around implementing it.

I have no idea where to start with open source/open weight models to make an on prem version viable and keep it updated.

@omarsar0 Been saying this for a few years

@random_walker Agreed, and the "learns over time" feature is exactly the trap. Every week it runs, more of your operational context lives inside their system and nowhere else... Hallucinations are inevetable 😅