Slack moved AI coding work out of the IDE and into a channel your colleagues can read

Slack launched Slack Code on August 20. Tag an AI coding agent on a piece of work and a channel opens around that task — the agent works inside it, the conversation that prompted the job comes with it, and the channel archives itself when the work is done.

Nothing in the mechanics is new; the placement is. Agent coding has mostly been a private exchange between one developer and one model inside an editor, reviewed by everyone else only once it surfaced as a pull request. Putting the plan, the diffs and the live previews in a channel means the intermediate steps are legible to people who were never going to open the editor, and Slack has added a step where a high-stakes change can route to a person for approval before it moves.

The consequence worth watching is auditability. Agent output now arrives as a channel artifact with permissions and a log already attached, which Slack points out means IT is not configuring or auditing anything new. That is roughly the shape a compliance function can accept. Getting an agent to produce a good diff and getting an organisation to accept one are separate bottlenecks, and this addresses the second.

Slack’s evidence for the workflow is its own, from its own team: over 70% of code channels, it says, open and close within a single day, from idea to merged PR. Slack has not published the denominator, the period measured, or the scope of what was counted. And the metric tracks how long a channel stays open, which is a fact about a channel rather than about the code — a task declared finished inside a day is not a task done well.

Where Slack put it is the other signal. The agent sits in the room where the team already talks, and Slack wired four vendors’ agents into it, Claude, Devin, Copilot and Vercel, on the bet that Slack remains where the work is discussed. That is a wager on distribution rather than on model quality, and it is the same instinct behind Microsoft plugging Copilot into the UAE’s sovereign AI stack and Meta turning its messaging apps into an agent surface: own the place people already are, and let the models compete inside it.

So what: the crowded lane is building coding agents; the open one is the reviewing, logging and approval layer around what they produce, which every regulated buyer will need before it lets an agent near production.

Related

• Microsoft’s Copilot is plugging into the UAE’s sovereign AI
• Meta is turning every WhatsApp and Instagram DM into a storefront
• ADNOC put AI across its whole rig fleet, and the critical data never leaves the country