Slack Code: bring your own agent, keep the review loop

On August 20, Salesforce shipped Slack Code: tag a coding agent in any Slack conversation and it spins up a dedicated channel where the work happens in public — diffs, live previews, a running plan, and a human who has to say yes before anything ships. Every AI coding tool announcement this year has been about making the agent faster. This one is about making the agent's work reviewable by people who aren't the person who prompted it.

What actually happened

Slack Code turns a coding task into a channel. Someone describes a bug or a feature, tags an agent, and the agent opens a project-specific code channel with tabs for the plan, the code changes, and a preview of the output. The team watches, steers, and audits diffs as they land. When the job is done the channel self-archives, leaving a searchable record of what the agent did and who approved it.

Launch partners named across VentureBeat's coverage and SiliconANGLE's include Anthropic's Claude Code, Cognition's Devin, GitHub Copilot, and Vercel's agent, with more reachable through API connectors. It runs on any Slack plan, including free workspaces — but you bring your own agent subscriptions. Slack sells the room, not the labor.

The permission model is the part worth reading twice: the agent cannot deploy to production on its own. High-risk actions need explicit sign-off from a person in the channel. Slack CMO Ryan Gavin framed the pitch against terminal-native tools: "All this individual work is a solo player sport. The knowledge and speed acting as an individual stays with you." Slack says it will open the code channel APIs to any developer, and that marketing and legal review are the next use cases.

Why agent review loops matter for your business

If you have one developer running Claude Code in a terminal, your review process is that developer's judgment at 11pm. That works until it doesn't. The failure mode isn't a bad agent — it's an agent that did exactly what one person asked, and nobody else saw the diff until it was in production.

We've watched small teams triple their pull request volume with coding agents and then discover the bottleneck moved. Generating code got cheap. Reviewing it did not. LinearB's 2026 benchmarks, drawn from millions of pull requests, found AI-authored PRs wait far longer for review and get accepted at a much lower rate than human-written ones. More output into the same review pipe is not more shipped software.

Slack Code is one answer to that. It is not the only one, and you should not need Slack to get it. What you actually need is three things, and you can build them into whatever you already run:

The agent's work is visible before it merges. Not a summary — the diff, in a place where more than one person will see it.

A human gate on anything irreversible. Deploys, migrations, customer emails, refunds. The agent proposes; a named person approves.

A durable record of who approved what. Six months from now, when something breaks, "the AI wrote it" is not an answer. "Dana approved this diff on August 12, here's the channel" is.

The bring-your-own-agent part matters too. Slack Code works with four different vendors' agents at launch. If your workflow is welded to one vendor's harness, you're one pricing change away from a migration. Build the review loop as yours, and swap the agent underneath it.

Sources: Salesforce, VentureBeat, SiliconANGLE.