Slack Code puts five coding agents in a shared channel — the part worth arguing about is where approval happens

Slack launched Slack Code on August 20, 2026. Tag a coding agent into a conversation and it opens a channel for that one task, with four tabs: the conversation, the agent’s plan, line-by-line diffs, and a live preview of the thing actually running. Anyone in the channel can pause, redirect, or stop the agent while it works. Nothing reaches production until a person signs off.

Five agents at launch — Claude, ChatGPT, Devin, GitHub Copilot, Vercel — on every Slack plan including free.

The coverage has mostly been about visibility: coding moves out of a private terminal and into the group chat. That’s the pitch, and it’s a reasonable one. But it isn’t the interesting part. The interesting part is that the approval gate has been moved into a Slack channel, and Slack channel membership was never designed to answer the question “who is allowed to ship this?”

The 30-second version

• What it is: a per-task Slack channel where an agent works in the open, with tabs for conversation, plan, diffs, and a live preview.
• What it costs: the channel is free on every plan. The agents are not — you bring your own Claude, ChatGPT, Devin, Copilot, or Vercel access. “Free on every plan” describes the room, not the worker.
• It is not Claude Tag. Claude Tag is Anthropic’s agent in Slack, Enterprise/Team beta. Slack Code is Slack’s surface hosting several vendors’ agents, all plans. Anthropic’s own framing is that they stack.
• The safety story is real but narrow: anyone in the channel can stop an agent mid-task. Good. The unanswered half is who counts as an approver.
• Our read: worth turning on for the visibility. Do not let it become the only review a change gets.

What it actually is

The per-task channel is the good idea here. A channel that exists for one piece of work, holds the plan and the diff and the preview together, and archives itself into something searchable is a better artifact than a thread that scrolls away. The live preview tab in particular does something a pull request doesn’t: it shows the thing running, which is how non-engineers actually evaluate whether a change is right.

Slack’s Rob Seaman, its EVP and general manager, put the case for it as AI only creating value when it’s part of how a team already works rather than something people go do alone in another tab. Anthropic’s Cat Wu, head of Claude product, described engineering work that already starts as a Slack conversation with Claude Tag and now gets a dedicated channel where the people, the context, and Claude work together from the start.

It is not Claude Tag, and the difference matters

Two products put a coding agent in Slack two months apart. They are not the same layer, and if you’re evaluating one you should know where the other sits.

One is an agent. The other is the room. Anthropic’s framing is explicitly that they compose — a conversation that starts with Claude Tag continues in a code channel. If you already run Claude Tag, Slack Code doesn’t replace it; it gives the work a container. If you don’t, Slack Code is reachable on a free workspace and Claude Tag isn’t, which for most teams is the more relevant fact.

The part the coverage skipped

“A person in the channel approves before code ships” is a sentence that sounds like a safety guarantee and is actually a permission model. Read it again as one.

Slack channel membership is loose by design. People are added to channels for awareness — a manager wants visibility, a designer needs to see the preview, someone from support is pulled in because they filed the bug. None of that is a statement about who should be able to green-light a production change. Repository write access is a deliberate, audited list. Channel membership is an ambient one.

Those two lists will not match, and the gap runs both ways. Someone with the authority to approve a change may not be in the channel. Someone in the channel for context may find themselves holding a button they weren’t meant to hold. Slack has not published the details of how approval authority is scoped, and we haven’t run Slack Code ourselves — this is a reading of the announced design, not a test result. But it’s the first question a platform team should ask, and it wasn’t in any of the launch coverage we read.

The mid-task controls are the better-designed half. Letting anyone in the channel pause or stop a running agent is exactly right: an agent heading somewhere wrong should be stoppable by whoever notices first, and making that the widest permission in the system costs nothing. It’s the shipping decision that needs the narrow list, and stopping that needs the wide one. Slack got the wide one right.

Does this replace the pull request

Slack isn’t claiming it does, and you shouldn’t act as if it does.

A diff in a channel tab is good for catching a wrong direction in the first ten minutes, which is when catching it is cheapest. What it doesn’t carry is everything review accumulates around a commit: the permanent record attached to the change, CI results, blame history for whoever reads this code in a year, and branch protection rules that encode who may approve what. Channels archive. Commits don’t.

The honest framing is that Slack Code adds a gate earlier in the process rather than removing the one at the end. Teams that treat it as an addition will get the benefit — problems caught before a PR exists, and people who’d never open GitHub able to say “that’s not what I meant” while it still costs nothing to change. Teams that treat it as a replacement will find out what they gave up during the first incident review, when the record of who approved what is in an archived channel instead of on the commit.

That’s the same tension covered in what happens when agents get destructive permissions: the fix is rarely a smarter model, it’s being deliberate about which action needs which approval.

What we’d want to know before turning it on

• Who can be in a code channel, and does that list match branch protection? Decide this before the first channel, not after.
• Does approval in the channel satisfy your existing review requirement, or is it in addition? If your compliance story depends on a recorded PR approval, an archived Slack channel is not the same artifact.
• What does the agent have access to? Slack Code is the surface; the credentials and repository scope belong to the agent you connected. That permission question doesn’t change because the work is now visible.
• What happens to the channel? It archives and stays searchable, which is good for reconstructing intent later — provided your Slack retention settings keep it as long as your incident reviews need it.

None of that argues against switching it on. The visibility is genuinely valuable, the stop button is well-placed, and the cost of trying it is a free channel. Just don’t let a well-designed room quietly become the place where the shipping decision gets made without anyone deciding that it should.

Companion reading

• Claude Tag, explained — Anthropic’s own Slack agent, who can use it, and the 65% claim worth reading twice.
• AI coding agents keep wiping production databases — why the permission model, not the model, is the fix.
• Best AI coding agents 2026 — where the five Slack Code partners actually stand against each other.

Sources

1. Salesforce Launches Slack Code to Make AI Software Development Multiplayer — Salesforce newsroom
2. Slack launches Slack Code, where teams and AI agents build together — The Next Web
3. Salesforce introduces Slack Code to bring agentic team coding into the open — SiliconANGLE
4. Slack Code Puts AI Coding Agents in Dedicated Project Channels — Unite.AI
5. Slack Code: AI Coding Agents Get Shared Channels for Review and Oversight — TechRepublic
6. Slack wants to drag AI coding out of the terminal and into the group chat — VentureBeat

FAQ

What is Slack Code? It launched on August 20, 2026. You tag a coding agent into a Slack conversation and it opens a dedicated channel for that task, with four tabs: the conversation, the agent’s plan, line-by-line code diffs, and a live preview of the running output. Anyone in the channel can pause, redirect, or stop the agent mid-task, and a person has to sign off before changes reach production. Launch partners are Claude, ChatGPT, Devin, GitHub Copilot, and Vercel.

Is Slack Code free? The channel surface is, on every Slack plan including free workspaces, from launch day. The agents are not included — you bring your own access to Claude, ChatGPT, Devin, Copilot, or Vercel, and that’s the part with a bill attached.

How is it different from Claude Tag? Different layers. Claude Tag is Anthropic’s agent living in Slack as a shared teammate, launched June 23, 2026 as a beta for Claude Enterprise and Team. Slack Code is Slack’s own surface hosting several vendors’ agents, available everywhere. Anthropic’s Cat Wu framed them as stacking: work that starts as a Slack conversation with Claude Tag continues in a dedicated code channel. One is an agent, the other is the room.

Does Slack Code replace pull request review? Slack isn’t claiming it does. A diff tab is good for catching a wrong direction early; it doesn’t carry the permanent record attached to the commit, CI status, blame history, or the branch protection rules that decide who may approve. Read it as an earlier gate, not a replacement for the last one.

Who can approve code in a Slack Code channel? Slack’s description is that a person in the channel approves, and that anyone in the channel can pause or stop an agent mid-task. That makes channel membership the control surface — and channel membership is normally managed loosely, for visibility rather than authority. It won’t automatically match your repository’s write access list. Settle who belongs in a code channel before you open the first one.

Which agents work with it? Claude (Anthropic), ChatGPT (OpenAI), Devin (Cognition), GitHub Copilot, and Vercel at launch. Salesforce separately named OpenAI, Google, Perplexity, Writer, Dropbox, Notion, Cognition, Vercel, and Cursor as bringing agents into the wider Slack ecosystem, which it puts at more than 2,600 third-party apps.