Looks like Anthropic is progressing further into platform territory and conquering Agentic use cases left right and centre. If you’re building an agent platform for workforce productivity today, your best best is model agnosticism and focus on token cost control.
It was very useful to delegate things like create a ticket based on a discussion to Claude.
The moat is to be anti lock in. The open source model needs to be amped up on steroids to avoid this permiating every facet of your org you cannot escape.
It recently wrote an entire dissertation for an epic, assuming it was related to some other project, where it had earlier made the wrong guess about a vendor capability (from their marketing materials, no less), and it all had to be thrown away. I cleared the memory, but it appears to be still pulling from some corporate data source i cant control or locate.
The only way it works is if customers truly start treating agents as humans with the same liability as an employee.
What's cooler is then it can view/add/remove people from channels, so it can conduct access reviews -- overall I consider it a security improvement.
But people can be invited to a channel after @Claude is provisioned. So yeah, I suppose you'll need to be deliberate about channel memberships.
Yeah, that explains a lot.
> @Claude is multiplayer. Within a given Slack channel, there’s one Claude that interacts with everyone. This means that anyone can see what it’s working on, and can pick up the conversation from where the last person left off. This makes tagging Claude very different from working within a single chat or for a single task—it’s much more like interacting collaboratively with a teammate.
In private context, I want to have a per-person conversation with durable context for that person's private chats. I also want that person's permissions to extend. Like contractors in our slack should only be able to ask and get back information about clients that they've been attached to, not our entire knowledgeable.
And we've implemented ALL that, but just with a lot of custom code. We've put in interceptors that put in per-user keys into the MCP connection so only certain tools are even exposed, etc.
[0] "Editing this message will create a new conversation branch. You can switch between branches using the arrow navigation buttons."
Anyway, if that happens, you've got organisational issues. Claude can't fix that.
From reading that and materials on it, it seems unclear if – let's say you do what's done in the demos on the site and 'dispatch work' from a thread in a shared channel (e.g. from some discussion) – that if any one of your coworkers replies below you and says, "Actually, could you fold in <blah> as well?" that Claude wouldn't listen to them and thus derail the work.
We do signal to Claude that there's a difference between a conversation's initiator versus incoming participants and we've found that in situations where people disagree on an approach, Claude patiently waits for a resolution while correcting any misunderstandings.
It's also worth mentioning that since Claude has its own identity, a coworker cannot enter a thread and commandeer _your_ identity; you collectively steer how Claude acts with its _own_ identity (it opens PRs as itself, browses Datadog as itself, etc).
Discussing what needs to be done next, and having it automatically sort it into what subtask it is, if it applies or is blocking to multiple other tasks. Recording specifications, measurements, dimensions. Being able to ask other people facts and have them correctly documented into the right task.
Company Brain / Knowledgebase is imho more rear facing, whereas todo is future facing.
Given the reliability and general product quality of the Anthropic product team's code, this doesn't sound like a selling point.
It takes multiple seconds to launch, random lines disappear in the scroll back, it’s internal state gets messed up causing the TUI to show duplicate and/or offset lines, and it frequently causes some kind of GPU buffer corruption causing the entire terminal env to show garbage.
Yes, this is the ubiquitous memory issue that I mentioned. Unfortunately, it is now the baseline in all modern apps.
>random lines disappear in the scroll back, it’s internal state gets messed up causing the TUI to show duplicate and/or offset lines,
I haven't seen this issue, other than when I am using a shell that is bugged and does it with all TUI/console programs (usually a virtualized shell which I resized). Do you have a reproducible example of this?
The trick is not bringing React into the terminal.
(FWIW, I have a link to a TUI harness in my profile that uses 50MB of ram and about 1% CPU while streaming, even in giant contexts)
Tons of examples of innocuous strings setting it off and sometimes with financial impact like the OpenClaw/hermes thing (just having the word "hermes" would insta deplete your quota and start charging you API rates in extra usage)
i.e.
- design is only available in web
- cowork is only available in desktop, sharing projects only works in chat, not cowork, which is arguably the more valuable place to have that feature (re: multiplayer like tag). SSO only works if you type your email, the "login with google" button be damned, and only after you finish typing your email does the login button text change.
- cli has a number of features only available there, with the cowork equivalent having a different name iirc
If you admin/support other people using the breadth of tools, you will see more of the slop they sling
Or possibly: they’re focusing where it matters?
Like okay Jan lmao, seems like the North Koreans take their uptime a little more seriously than the nuclear weapons developers at Anthropic.
“See, here is our company made to worth X by our product, because it can make a company worth X.”
And yet, the product made something really unreliable.
And nobody bats an eye on the market.
But hey, SpaceX successfully sold that they would generate 30 trillion dollars of revenue.
edit: note this may not be official release, and may be unavailable for some users. I saw it show up yesterday listed as available Mcp and used it to view projects in Claude design.
For example, it writes the whole front-end twice, once is claude design and then later it has to read it again and re-implement it in code. Also a lot of stuff (e.g. Claude.md, skill files, etc) are not supported, and they have their own set of ui-design and design systems, which claude code doesn't support.
I think Claude Design is a wonderful product, i'm just pointing that it is an independent product to Claude Code, heck even today it works better with Codex than Claude Code (that is how i use it with my own browser-use agent, i tell it to browse the design in Claude.ai/design and re-implement the design, works much better than downloading the zip file and asking the model to implement in that way.
The spooks at Langley aren't going to produce such a file for run-of-the-mill encounters that you have with the legal system, but Anthropic will.
I know my users would actually like Claude Tag, but unfortunately we are in Teams, not Slack, as are most other non-tech companies.
Cowork/Claude Desktop itself is also quite a frustrating product. There's no native audit log unless you basically wire up your own with the API & a log aggregator. You can't selectively enable Claude Code access per team member, it's all or nothing. Some of the MCP connections (like QuickBooks Online) don't do RBAC at all, it's all or nothing for every user in the team.
Maybe enterprise isn't their target market, but they do keep making features that make it seem like they want that market. But if they do, they really need to step up work on governance features & RBAC for specific features and settings per member in the team.
If they don't, Microsoft will eat their lunch for enterprise non-programming use.
https://www.microsoft.com/en-us/microsoft-365/blog/2026/03/0...
I expect many of those to be shut down sooner than later, if learning anything from Google.
@claude collect all the internal knowledge and context. And fire folks who are not required.
Would also be funny if it fires the one who done the request.
This announcement is all about Claude extending its reach beyond single-player workflows and into multi-player workflows.
On the flip side, Slack just announced MCP support for the Slackbot AI chat capability embedded within Slack. It is, for now, exclusively single-player.
Single-player is the "safe place" (relatively speaking). The context, permissions, and standards (MCP/MCP UI apps) all work reasonably well for it, but get super complex or break down entirely when thrown into a multi-player shared context. I suspect Slack is doing what they are doing with an eye towards multi-player, but it's hard to say how that will manifest.
For a real life example of this challenge: I work in scheduling (for Reclaim.ai) and you can ask our chat to find time to meet with a coworker and we go find time and help explain why certain times won't work. For example, it might say: "I couldn't do 11am tomorrow because you've got a job interview scheduled on your personal calendar". This is safe and fine to do in a private context.
But... imagine if one were to ask our service (or Claude) to find time with someone and it replied to the thread for everyone to see: "The soonest I could find is 12pm tomorrow. Reggie is available at 11am, but Lightbody has a job interview so it won't work". WHOOPS.
I think the other comments in this thread have the right idea of it: for this to really work safely, the permissions model needs to be nailed down, and it may mean that you end up with multiple identities of "Claude Tag" (or whatever agent you engage with in a public forum), and the context it gets is only the context that particular identity is entitled to, just like any other employee. But then that gets tedious because now I've got even MORE "people" to keep track of and know who to engage with, which is half the problem getting work done in large enterprises.
Will be interesting to see how this evolves. I'll have my popcorn out :)
Work should be entirely separate calendars with things from personal added to it in generic blocks the owner can understand but protects from this leakage. Work calendars are owned by companies (generally) and should be treated like the are looked at by coworkers.
My point is: context and privacy is HARD in multi-player modes.
For enterprises already with an Anthropic MSA, hard to see the argument to purchase a third party, like Glean, over this.
The best part for me is seeing non-technical folk spec out something in a thread that they discussed something and letting the agent go ahead and build it ready for the humans to review later.
The difference between this and our agents is that they are context aware - i.e. you can use them privately to access personal information safely.
Can provide a link if interested.
During an incident, how do I know which Claude Tag called AWS?
> Think of it as creating separate Claude identities for different uses: everything, including its memories, will stay scoped to the channels defined by the administrators. For example, a model set up for sales work won’t pass on memories to one set up for engineering; nor will it give engineers access to any sales data or tools. More information about provisioning access is available here (https://claude.com/blog/agent-identity-access-model).
But they walk all this back by saying that for PRs in GitHub it uses the upstream Claude for GitHub app - that is one installation of a GH app, which means one identity, one list of repos it can access. In the audit logs it will be impossible to see if it was Claude Tag from Channel Y or Claude Tag from channel X.
Arguably that’s a limitation of the abysmal state of machine identities in GH.
Agent identity and attribution continues to be important, but current systems makes it oh so difficult.
AI enables quick shipping, but the traditional moat of development no longer applies.
Devin is model agnostic, and isn't Slack only.
They can't really comment the truth about it publicly or the rest of the company gets banned too.
The hard problem is giving a shared agent durable organizational memory and a real isolated environment where it can safely access company systems and perform work. The agents need a durable log of what everybody at the company is doing, prevent data leaks with proper access control and isolate the runtime to give everybody both private & shared space.
It’s also not tied to Claude or Slack. We see Slack as one interface and the models as part of the harness. It's usually better to combine multiple providers to review the work.
They shoudl just expose SDK/platform for people to build own integration with Discord/Teams/mattermost etc. This would allow for fine grained permisions control and speed up adoption. To large degree this is alternative for OpenClaw ?
People use Teams only because it's basically free when you get the M365 suite for your company, not because it's good or usable in any way.
https://x.com/EngramLab/status/2069465879696576844
https://x.com/karlmehta/status/2069329697843151333
Satya Nadella reveals why every company may need its own AI model: the model becomes the new company database.
"To me, a model is like the database market."
"A firm should be able to take the tacit knowledge it has and embed it inside weights in a model that they control."
"When somebody asks me how many models should there be, I'll say as many models as firms in the world."
The contrarian part: the value may not sit in one universal frontier model. It sits in each company turning its private operating knowledge into a controlled model.
More importantly, claude-slackbot automatically remembering sounds like it'll be company wide AI slop? 75% of the stuff on slack should not be remembered for the future. And 90% of the stuff worth having in context is not even on the slack thread.
A tiny detail...
Has everyone collectively lost their minds? Suggesting anything like this even five years ago would get you laughed out of the room, and actually doing this would be a career ending mistake.
That explains a lot.
Not sure how much we'll use this, but it could be useful for filing tickets from conversations. Though I'd prefer to just point claude to the convo post-hoc rather than have to invite it each time just in case I want to ask claude to do something.
The last thing I want is claude chipping into a convo Clippy-style.
Although I suppose the problem with doing it for Cowork is this is a slack plugin, and that is not where most non-tech companies are. Teams is, at 320+ million active users vs. ~50 million.
Everyone hates Teams, but like or not that is where enterprise work happens, not slack. Anthropic would do well to make their Microsoft 365/Teams integration story better and go after enterprise before OpenAI does, or before Microsoft catches up (if they ever do) with Copilot.
I think that people see the word "claude" and smash the upvote button. I don't think it's botted. My guess is that people just want another place where they can discuss ai coding workflows.
I know this is much easier said than done. And my comment is partly tongue-in-cheek.
But Slack isn't great. It isn't loved.
I don't understand why Slack still doesn't have a true single inbox view of all activity.
And their LLM integration is poor, for what feels like one of the clearest use cases for a rich LLM chat experience -- to discuss what's happening in a channel, filter cruft, and ask an agent to take actions against MCP productivity tools like add a todo, etc.
Slack's current AI integrations are piss poor. Just summaries with zero customization. I mean, isn't it obvious that's not good enough?
Eventually a ground-up rethink of Slack in an LLM age is going to displace Slack.
They'll likely survive forever as an enterprise provider, like Microsoft Office and related tools, simply because of integrations, etc.
But Slack feels ripe for disruption.
I'm still using it all the time and getting an immense benefit.