AI coding tools
Slack Code Brings Agent Workflows into Shared Channels, but the Custom Agent API Is Not Yet Widely Available
Slack Code lets Claude, Devin, Copilot, ChatGPT, and Vercel agents present plans, code diffs, and live previews in project-specific channels. It improves the interface for multi-person agent oversight, but execution sandboxes, credentials, and repository permissions remain the responsibility of each agent provider.

Salesforce has launched [Slack Code](https://www.salesforce.com/introducing-slack-code/), an effort to transform coding agents—previously confined to individual terminals or web sessions—into units of work that entire teams can observe and control. Users can mention a supported agent in an existing conversation, and Slack will create a “code channel” tied to a single project. The channel provides tabs for conversations, agent plans, code diffs, and executable HTML previews. Members can also pause, redirect, or terminate the agent. Once a task is complete, the channel is automatically archived, while its records remain available for search and audit purposes.
The initial integrations include Anthropic Claude, Cognition Devin, GitHub Copilot, ChatGPT, and Vercel agents. The key change is not the addition of another chat interface, but the consolidation of agent state, artifacts, and human approvals into a shared object. A product manager can initiate a task from a bug report, an engineer can then review the diff and preview, and the team can finally approve the creation of a PR or a deployment. Slack also says the code channel API can be used to automate account provisioning, authenticate users, and customize workflows, but an API for general developers has only been promised for a later date.
The security boundaries need to be clearly distinguished. Slack says agents inherit the workspace’s identities, permissions, and administrative policies, and that high-risk actions require human approval. However, the official materials do not explain what kind of sandbox actually executes the code, how network egress is restricted, how secrets are injected, or how repository authorization maps to channel members. Each partner agent still requires a separate account and access permissions, so enterprises should evaluate each provider’s data-retention, branch-protection, and credential models individually rather than treating the Slack channel itself as a complete security boundary.