Slack Code Brings Coding Agents into Shared Channels, With Diff Views, Live Previews, and Human Approval

Salesforce launched Slack Code on August 20 in an effort to move coding-agent workflows—previously confined to individual IDEs, terminals, or browser tabs—into Slack workspaces where entire teams can participate. Users can mention an agent in an existing conversation, prompting the system to create a Code Channel tied to a single project. After receiving requirements, screenshots, and documents from the channel, the agent can propose a plan, modify code, and display execution results.

A Code Channel is not a conventional chat thread. Its interface separates conversations, plans, line-by-line diffs, document Canvases, and live HTML previews into dedicated tabs, enabling product, design, and engineering teams to align on direction before the agent opens a pull request. When a task is complete, the channel is automatically archived, while its searchable conversations and artifacts are retained to create an audit trail spanning requirements, agent actions, and human approvals. Any member can also pause, redirect, or terminate the agent. High-risk actions, such as deploying to production, require an expert to approve them within the channel.

The initial group of supported agents includes Anthropic's Claude, Cognition's Devin, GitHub Copilot, ChatGPT, and Vercel agents; users must obtain access to the corresponding services separately. Salesforce says the Code Channel API can handle account provisioning, user authentication, and workflow automation, with support for integrating custom agents planned for the future. This moves Slack closer to serving as a coordination and approval control plane for agents, rather than merely a notification front end.

For now, engineering teams should not equate “inheriting Slack permissions” with comprehensive code-execution security. Salesforce has not yet detailed the specific mechanisms for agent sandboxing, repository permission mapping, secret injection, branch protection, or safeguards against malicious channel content. The question of who retains final control when multiple participants issue conflicting instructions also requires real-world testing. Key areas to watch next include the API specification, whether event logs can be exported, and whether all partner agents consistently comply with approval and revocation workflows.