Slack just gave AI coding agents a seat in the group chat

Marc Benioff announced Slack Code at Dreamforce last week, and his post laid out the whole pitch in fourteen words: "Don't code alone. Slack Code is live. Humans and agents. Same channel. Same work," wrote @Benioff, listing agents from Anthropic, GitHub, Cognition, and Vercel as launch partners. @SlackHQ followed with its own framing, that "AI work isn't a solo task anymore, it's multiplayer," and named Claude Code, Devin, GitHub Copilot, ChatGPT, and Vercel agents as the tools you can now pull into a channel. Cognition's Jeff Wang shipped the Devin integration the same day, opening his announcement with "A good teammate earns trust," posted @jeffwang.

What actually shipped

The mechanics are simple. You mention a coding agent in any Slack conversation and it spins up a dedicated code channel scoped to that task. From there:

• Everyone in the channel sees the same thread the agent is working from.
• Code changes surface as diffs the team can review, and HTML output gets a live preview.
• Feedback left in the channel gets folded back into the agent's work.
• Someone approves the result before it ships, the channel archives itself when the job is done, and an audit log sticks around.

It works on every Slack plan, including free ones. That detail matters. Salesforce is not selling this as an enterprise add-on. It wants agent collaboration to be a default behavior of Slack itself.

The feature is oversight, not coding

The easy joke landed within hours. Gizmodo's headline was "Slack Has (of Course) Launched a Vibe Coding Tool," and that skepticism is fair. Agents could already work through Slack bots, and a chat window is nobody's idea of a serious development environment.

But the coding is not the interesting part. The supervision is. For the past two years, agent work has mostly lived in one developer's terminal or one browser tab. The agent might write half a feature, and nobody else sees anything until a pull request appears. Every piece of Slack Code that isn't the agent, the shared thread, the visible diffs, the approval step, the audit log, exists to fix that. Benioff called it multiplayer coding. I'd call it witnessed coding, and witnesses are exactly what agent work has been missing.

It also answers an etiquette problem the AI world spent this month arguing about: agents that ping your coworkers directly are annoying, a point OpenAI's president conceded when ChatGPT joined Slack. A task-scoped channel that archives itself is a reasonable fix. The agent gets a room. It does not get your DMs.

The honest caveats: this pulls your development workflow deeper into Salesforce territory, "multiplayer" can degrade into five people watching a robot type, and an approval step only works if approvals stay real. If sign-off becomes a rubber stamp, the audit log records the failure instead of preventing it.

What to steal from this, even if you never code

Most business owners who say they don't trust AI mean something more specific: they can't see what it did before a customer does. Slack Code is one company's answer, and the pattern underneath it is portable to any business, no developers required. Put the AI's output where your team already talks. Make it reviewable before it goes out the door. Keep a record of who approved what.

An agent that drafts quotes, invoice reminders, or customer replies into a shared channel, where your office manager taps approve, is a very different risk than one that sends on its own. The first version is how adoption actually happens in a small company.

If you're weighing where that kind of checkpoint would fit in your own shop, New Face Design runs a free process audit for exactly this question: which tasks an agent should draft, and where the human sign-off belongs. The frontier labs are building the supervision layer for coders. The same idea works just as well for a five-person office in the Fox Valley.