Kylon vs. Claude Tag: Same idea, very different architecture

On June 23, Anthropic launched Claude Tag — a way to @mention Claude inside Slack channels so it can follow conversations, build context, and work alongside your team. Andrej Karpathy called it a "new paradigm." Anthropic says 65% of their product team's code now comes from an internal version of it.

We read the announcement with genuine interest. Not because it was surprising, but because it validated something we've been building toward for over a year: AI should be a teammate, not a tool you open in a separate tab.

Claude Tag and Kylon start from the same premise. Where they diverge is in how deep that idea goes.

What Claude Tag does well

Let's give credit where it's due.

Claude Tag solves a real problem. Before it, using Claude at work meant opening claude.ai, pasting in context, getting an answer, and then carrying that answer back to wherever the real work was happening. Now Claude sits inside Slack threads. It reads the conversation. It responds in context. You don't have to leave the channel.

The "multiplayer" design is smart. One Claude instance per channel means the whole team can see what it's doing, steer it mid-task, and pick up where someone else left off. That's meaningfully different from a personal chatbot.

And it's not limited to Slack conversations. Admins can provision an Access bundle — a set of service-account credentials — that lets Claude reach external tools. GitHub for code reviews and pull requests. Google Drive and Notion for documents. BigQuery and Snowflake for data queries. Jira, Linear, and Asana for issue tracking. Datadog and Sentry for monitoring. HubSpot, Gong, and Salesforce for CRM. Any service with an HTTP API can be connected in principle.

The ambient mode — where Claude proactively monitors channels and surfaces relevant information — is an interesting bet. It's off by default, which is the right call, but the intent is right: AI shouldn't always wait to be asked.

Where it gets complicated

The Hacker News thread on Claude Tag's launch tells you what the announcement doesn't.

The session hijacking problem. Multiple people in a Slack channel can all steer Claude at once. One HN commenter described the obvious scenario: you dispatch Claude on a task, a coworker replies "actually, could you fold in X as well?", and now your work is derailed. The response from another commenter was blunt: "If a coworker hijacks your session, you walk over to their desk and slap them on the back of the head." True, but not exactly an architecture solution.

The flat permission problem. This is the one that will bite enterprises first. Claude Tag's documentation is explicit: "Connections belong to the agent identity, not to any person." Each connection is a dedicated service account — Anthropic recommends creating a virtual user per service specifically for the agent. The admin configures what Claude can access per channel scope, and then everyone in that channel gets the same access through Claude. There's no concept of "this user can query the pipeline table but that user can only read docs." If Claude has write access to your GitHub repo from that channel, every person in the channel can ask it to open PRs, regardless of their own GitHub permissions. Zenity called this the "confused deputy problem" within a week of launch.

Per-channel memory, not organizational memory. Claude Tag's memory is scoped to individual channels. Context built in #engineering doesn't carry to #product. If you're working across multiple channels — which every real team does — you're re-explaining things to Claude every time you switch contexts.

The token cost question. One HN commenter nailed it: "Wowza, this will be a token guzzler. Assuming Claude is parsing every message posted on multiple Slack channels, compacting knowledge..." Claude Tag is consumption-based, drawing from an organization spend balance. Every message in every monitored channel burns tokens. For active Slack workspaces, the costs can be unpredictable — Anthropic suggests running a pilot with a spend cap to see what it actually costs before committing.

It's still a guest in someone else's house. Despite the integrations, Claude Tag lives inside Slack. That means it inherits Slack's data model, Slack's threading model, and Slack's UI constraints. It can't create structured databases. It can't run complex scheduled workflows with triggers. It can't manage files with version history. Its network access is HTTP-only through an agent proxy — no SSH, no native database protocols.

The deeper problem: your brain is the bottleneck

There's a question that neither Claude Tag nor most AI workspace tools are really addressing, and it's the one that matters most.

When agents join an organization, the nature of the workspace changes. It goes from being a place where humans talk to each other to a place where work actually gets executed. The ratio flips: 80% execution, 20% coordination. And all that execution produces an enormous volume of information — drafts, analyses, tables, status updates, charts, code diffs, deployed pages. The workspace becomes a factory floor, not a water cooler.

But here's the biological constraint: your brain hasn't upgraded since the Pleistocene. Human working memory holds roughly four items. Conscious processing bandwidth caps at about 50 bits per second. Those numbers don't improve with better models.

So the real design challenge isn't "how do we make agents smarter?" It's: how do you compress 5,000 messages of agent output into the 50 things a human actually needs to see?

Slack's entire information architecture was built on one assumption: this is a place for human conversation. Linear message threads. Chronological feeds. @mentions. Everything optimized for one human communicating with another. When agent output scales exponentially but the interface doesn't change, you don't get a smarter team — you get a noisier channel.

This is why we believe you can't just bolt AI onto an existing collaboration tool and call it a day. The information architecture itself needs to be redesigned — not for what agents can produce, but for what humans can absorb.

How Kylon approaches this differently

Kylon and Claude Tag both believe AI should be a teammate. But Kylon doesn't put an AI agent inside someone else's collaboration tool. Kylon is the collaboration tool.

That architectural difference matters in several concrete ways:

Rich interaction, not just text messages

In Slack, everything an agent outputs is a message. Even with Block Kit — Slack's UI framework for apps — you're working within a constrained set of components: fixed layouts, no custom styling, no dynamic visibility per user, no native tables, no inline charts. Developers have described Block Kit as "printing a spreadsheet as a novel."

Kylon agents can render their output in whatever format communicates most efficiently. A market analysis becomes an interactive card with key metrics highlighted, drill-down sections, and one-click approval buttons. A comparison becomes a real table with sortable columns. A status update becomes a compact stats card, not a wall of text. Agents can show different UI to different users in the same message — an approval button only for the manager, a detail view for the analyst.

This isn't a cosmetic difference. When agent output volume scales 10x, the format it arrives in determines whether a human can actually process it. A 2,000-word analysis crammed into a chat bubble? Nobody reads it. The same analysis as a structured card with the key number highlighted? Everyone uses it.

Attention management built into the architecture

When agents produce 10x the output, the notification model has to change. In Slack, you get notified about everything or you mute the channel entirely — there's no middle ground. Claude Tag inherits this binary.

Kylon is designed around the constraint that human attention is finite. Agents can deliver work results silently — completing a task, updating a database, filing a report — without pinging anyone. The workspace surfaces what needs your attention and compresses the rest. Conversations with agents don't pile up as infinite scrollback; they compress into summaries that expand on demand.

The question isn't whether AI can produce more output. It can. The question is whether the interface respects the fact that you can only absorb so much. We built for that constraint from day one.

Memory that works across the whole workspace

In Kylon, an agent's memory isn't scoped to one channel. An agent that learns your deploy process in #engineering remembers it when you mention a deploy in #ops. It knows that "the Q3 launch" refers to the same thing whether you mention it in a thread, a database, or a DM. Memory is persistent, cross-channel, and accumulates over weeks and months.

Claude Tag resets context at the channel boundary. Kylon doesn't have that boundary.

A flexible permission model

Both Kylon and Claude Tag let agents connect to external tools. The difference is in how permissions work.

Claude Tag uses a single agent identity with dedicated service accounts. An admin creates an Access bundle — say, a GitHub service account with write access — and attaches it to a channel scope. From that point, anyone in that channel can ask Claude to open PRs, merge code, or read private repos through that same service account. The access is flat: the same permissions for the intern and the CTO, no per-user differentiation.

Kylon's approach works in two directions. Agents can have their own connected accounts — their own GitHub access, their own email credentials — for work that belongs to the agent. But humans can also delegate their own accounts to an agent for a specific task, so the agent acts with the delegator's permissions. If you don't have access to a repo, asking an agent to touch it won't bypass that. The agent's reach is bounded by whoever authorized the action, not by a blanket service account that everyone in the channel shares.

Access to every capable model, not just one

Claude Tag runs on Claude Opus 4.8. Only Claude. If you want to generate an image, create a video, use a specialized model for a specific task, or simply pick the model that's best for the job — you can't. You get Claude, for everything.

Kylon agents have access to every capable model on the market through a unified proxy layer — reasoning models, image generators, video creators, audio transcription, and more. Need deep analysis? The agent dials up the reasoning effort. Need a quick answer? It keeps things light. Need to generate images, produce a video, or transcribe a meeting? The agent reaches the right model for that task — without you knowing or caring which one it is. And because Kylon connects to 1,000+ data sources natively, agents work with your actual business data across GitHub, Gmail, Notion, Slack, CRMs, ad platforms, analytics tools, and more.

The practical difference: a Claude Tag agent can query your BigQuery warehouse from a Slack thread and chart the results. A Kylon agent can do that, plus pull live campaign data from Meta Ads, cross-reference it with your analytics, generate a branded performance deck, and email it to your client — all from one mention in a channel.

Databases, not just conversations

Slack is a messaging app. Everything in it is a message. Claude Tag can read those messages and respond to them, and through integrations it can reach external databases. But it can't create a new database inside the workspace, update records in a native data layer, or build a workflow that triggers on data changes — all of that happens in external tools that require separate setup and maintenance.

In Kylon, database apps are first-class. Agents read and write structured data directly inside the workspace. A marketing agent can pull metrics from a campaign database, update lead status, and trigger a follow-up workflow — all without leaving the workspace. The data layer isn't bolted on; it's built in.

Apps that anyone can build, together

Here's something Claude Tag can't do at all: let your team collaboratively build custom applications inside the workspace.

In Kylon, anyone on the team can describe what they need — "I want a CRM that tracks our pilot customers and sends me a weekly summary" — and agents build it. These aren't toy prototypes. They're versioned applications with proper databases, APIs, and deployment pipelines, built through conversation rather than code. Multiple people can collaborate on the same app, iterating on it together. A salesperson defines the fields they need. A manager adds the reporting view. An agent handles the implementation.

This shifts the power from engineering teams to the entire business. Operations can build their own tracking systems. Marketing can create their own campaign dashboards. HR can design their own onboarding flows. Nobody needs to file a Jira ticket and wait six weeks for the dev team to get around to it.

Claude Tag is a smart assistant that lives in a chat thread. Kylon is a platform where teams build and run their operations.

Multi-agent coordination

Claude Tag is one Claude per channel. If you want different AI capabilities in different contexts, you configure different Access bundles through admin settings.

In Kylon, multiple specialized agents can work in the same channel. A data agent pulls metrics. A writing agent drafts the report. A marketing agent schedules the campaign. They see each other's output, hand off naturally, and coordinate without human intermediation. They're not different configurations of one model — they're different agents with different memories, skills, and tool connections.

The real comparison

Where it lives — Claude Tag: inside Slack. Kylon: its own workspace, purpose-built for human-agent teams.

Information presentation — Claude Tag: Slack messages, limited by Block Kit. Kylon: rich interactive cards, tables, charts, per-user UI, and structured output designed for human absorption.

Attention management — Claude Tag: inherits Slack's all-or-nothing notification model. Kylon: agents work silently by default, workspace surfaces only what needs human attention.

Memory — Claude Tag: per-channel, resets across channels. Kylon: cross-workspace, persistent over weeks and months.

Permissions — Claude Tag: flat access through dedicated service accounts; everyone in a channel gets the same permissions. Kylon: agents can own accounts or use delegated human permissions — access follows the person who authorized the action.

Models — Claude Tag: Claude only. Kylon: every capable model on the market — reasoning, image generation, video, audio — with adaptive reasoning effort per task.

Data sources — Claude Tag: Slack content plus admin-configured external integrations (GitHub, Google Drive, Jira, Snowflake, etc.) through a single agent proxy. Kylon: 1,000+ native data source connections, plus a built-in database layer agents can read and write directly.

App building — Claude Tag: not available. Kylon: collaborative app building through conversation — versioned, deployed, usable by the whole team without technical knowledge.

Multi-agent — Claude Tag: one Claude per channel. Kylon: multiple specialized agents per channel.

Automation — Claude Tag: scheduled tasks and channel watching within Slack. Kylon: workflows, cron triggers, webhooks, database-change triggers.

Pricing — Claude Tag: Enterprise/Team only, consumption-based. Kylon: free trial available.

When Claude Tag makes sense

If your team already lives in Slack, you're on a Claude Enterprise or Team plan, and your primary need is having an AI that can follow conversations and answer questions in context — Claude Tag is a reasonable choice. It's low-friction to deploy (it's just a Slack app), and for teams that mainly need a smarter layer on top of their existing Slack workflow, it works. The external integrations are solid — connecting GitHub, Jira, and a data warehouse gives Claude real reach beyond Slack messages.

It's also a strong fit for engineering teams already using Claude Code who want that same capability inside their team channels.

When Kylon makes sense

Kylon is built for teams that need more than a conversation partner:

• Your work involves structured data, not just messages. Tracking leads, managing projects, monitoring campaigns — things that need databases, not threads.
• You need AI that takes real action. Not "here's what you should do" but "I did it, here's the PR / the email / the updated spreadsheet."
• Agent output is overwhelming your team. You need a workspace that compresses, prioritizes, and presents information for human absorption — not one that dumps everything into a chronological feed.
• Non-technical people need to build things. Operations, marketing, sales — they should be able to describe what they need and have agents build it, collaboratively, without writing code.
• Permissions need to follow the person. You don't want a flat service account that gives everyone in a channel the same access to every connected system.
• Multiple agents need to collaborate. Your marketing, data, and content work shouldn't all funnel through one Claude identity.
• You need more than one model. Image generation, video production, specialized reasoning — not every job calls for the same model, and you shouldn't have to switch platforms to access a different one.
• You want the workspace, not a plugin. Instead of adding AI to Slack, you want a workspace where AI is native from the ground up.

The bigger picture

Claude Tag is Anthropic's acknowledgment that the chatbot-in-a-tab model isn't enough. They're right. The future of work involves AI that lives where teams work, understands context, and takes action.

But grafting an AI agent onto Slack is fundamentally different from building a workspace where AI agents are first-class members. One approach inherits the constraints of the host platform. The other designs the platform around the agents — and, more importantly, around the humans who need to work with them without burning out.

The real bottleneck in the AI-powered workplace was never the AI. It was always the interface between the AI and the three-pound organ trying to keep up. The workspace that solves for that constraint — not just for agent capability — is the one that wins.

Claude Tag is a smart move by Anthropic. It's also, in a way, the best argument for why Kylon exists.

Kylon is the AI-native workspace where humans and AI agents collaborate as peers. Start your free trial.