Slack Code launches — AI coding agents now work in shared Slack channels

What happened

On August 20, Salesforce launched [Slack Code](https://www.salesforce.com/introducing-slack-code/), a feature that puts AI coding agents into dedicated, shared Slack channels instead of a single engineer's terminal. Tag in an agent — Claude Code, GitHub Copilot, Cognition's Devin, ChatGPT, or Vercel's agent are supported at launch — and it spins up a "code channel" scoped to the task: fix a bug, update a page, ship a feature. Everyone added to that channel watches the same conversation, reviews the same code diffs, and checks the same live preview as the agent works. It's live now on every Slack plan, including free, at no extra cost, and confirmed by [The Register](https://www.theregister.com/saas/2026/08/20/slack-code-taps-into-collective-vibe-puts-ai-agents-into-the-group-chat/5290413) and [Salesforce Ben](https://www.salesforceben.com/salesforce-brings-vibe-coding-to-slack-with-new-slack-code/).

What's genuinely new

Agentic coding tools aren't new — Claude Code, Devin, and Copilot have all been writing and shipping code unsupervised for a while now. What's new is where the oversight lives. Previously, if an agent went off the rails, the only person who'd notice was whoever was staring at that terminal. Slack Code moves the entire loop — prompt, diff, preview, questions, approval — into a channel that a product manager, a designer, or a second engineer can see without being looped in separately. For anything that touches production, the agent packages its changes and waits for an explicit human sign-off inside the channel before it ships. Salesforce says agents inherit Slack's existing permissions and admin controls automatically, so IT doesn't have to provision a separate approval system for each tool.

That's the actual shift: not "agents can code," which we already knew, but "the record of what an agent did and who approved it now lives in the same place your team already works, by default."

What it means for a business owner

Most business owners running AI agents — coding or otherwise — have the same blind spot: the agent's output is trusted because nobody but the person who deployed it ever looks at the trail between the instruction and the result. Slack Code is a specific bet that shared visibility plus a mandatory approval gate before anything irreversible happens is the right default for agentic work, not an optional add-on bolted on later.

That pattern generalizes past code. If you're running an agent that drafts customer emails, updates a CRM record, or reconciles an invoice, the question worth asking isn't "is the agent capable enough" — most of the mainstream ones are, for scoped tasks. It's "who else besides the person who set it up can see what it did, and is there a real approval step before the consequential action, or does the agent just act?" Slack Code makes that visible-by-default pattern free and built-in for coding teams. For everyone else, it's still something you have to design in deliberately — a logged run history, a named approver, a step that blocks before the send/write/charge action.

If your team is already on Slack and has developers experimenting with Claude Code or Copilot, this is a low-cost way to make that work visible to non-engineers, which is often the actual barrier to trusting agent output enough to rely on it.

The honest caveat

This doesn't solve the underlying trust problem, it just gives it a better UI. Visibility isn't the same as verification — a product manager watching a diff scroll by in a Slack channel is not the same as someone who actually reviewed the security implications of what shipped. Salesforce's own description leans on "keeps people in control without slowing things down with separate IT approvals," which is a marketing framing for "we didn't build a new approval system, we reused the one you already have" — Slack's existing permission model wasn't designed with agent-generated production code in mind, and treating it as sufficient governance is a judgment call each team has to make, not a guarantee the feature provides. If your team doesn't already have someone with the authority and the actual technical judgment to gate a production push, moving the conversation into Slack doesn't create that person.

What to do about it

If you're running or considering AI agents for any business process — not just code — don't wait for a vendor to build you a shared-visibility feature. Write down, for each agent task in production, who can see what it did and what has to happen before its output becomes irreversible. If the honest answer is "only the person who set it up, and nothing," that's the gap to close before you add more automation, regardless of whether Slack Code or anything like it ever touches your stack.