Slack launches Code with Anthropic, GitHub, Cognition and Vercel agents

Why it matters

Slack Code makes Salesforce the coordinator rather than the model provider, giving competing coding agents one team interface while Slack captures the context around their work.

Marc Benioff (@Benioff) launched Slack Code on August 20th, pitching a shared workspace where software teams can assign work to coding agents from Anthropic, GitHub, Cognition and Vercel without leaving a Slack channel.

https://x.com/Benioff/status/2090240159115853956

"Don't code alone," Benioff wrote in the launch post, calling the product "multiplayer coding." Slack Code is being demonstrated at Dreamforce 2026, according to Benioff, who said the four partner integrations were live on Thursday.

Benioff co-founded Salesforce in 1999 after 13 years at Oracle, then spent much of the next two decades moving business software from locally installed applications into the browser. Slack Code extends that playbook to AI development: the interface remains familiar, while the work moves to agents operating against repositories, issue trackers and deployment systems.

The launch also gives Salesforce a sharper developer story for Slack, which Salesforce acquired in 2021. Slack has spent the past year repositioning itself as the coordination layer for workplace agents rather than another place for employees to exchange messages.

A front door for agents already doing the work

Slack Code arrives after each partner had already built pieces of the workflow Benioff is now packaging under one name.

GitHub added its Copilot coding agent to Slack in October 2025. Engineers can mention the GitHub app in a thread, assign a bug fix, test task or refactor, and receive a link when Copilot has prepared a pull request. GitHub later added natural-language issue creation, allowing discussions in Slack to become structured repository work without a separate handoff.

Anthropic introduced Claude Tag in June 2026 as a shared Claude identity that teams can add to selected Slack channels and connect to tools, data and codebases. Unlike a private chatbot session, Claude Tag retains channel context and lets multiple employees follow or continue the same task. Anthropic said the beta initially targeted Claude Team and Enterprise customers.

Cognition has long treated Slack as a primary interface for Devin. According to Cognition's documentation, engineers can tag Devin when a bug report or feature request appears, answer its questions in the thread and wait for it to return with code or a pull request. Cognition explicitly warns customers to review Devin's output.

Vercel, meanwhile, has been building the deployment side of this workflow. Its Slack agent tooling connects coding agents with Slack events, approval steps, execution environments and Vercel deployments. That gives Slack Code a route from a request discussed in chat to code that can be run and previewed, rather than stopping at an agent-generated answer.

Benioff's initial announcement does not establish that the four agents share one execution environment or a common model layer. The shared element is the Slack channel, where humans can see requests, agent replies and work status in the same conversation. Each provider still brings its own agent, account relationship and underlying development infrastructure.

Slack wants the conversation to become the control plane

Slack laid the groundwork at Dreamforce 2025, when it described itself as an "agentic operating system" and named Anthropic, Cognition and Vercel among the developers building agents for the platform. Slack's platform work included a real-time search interface, a Model Context Protocol server and developer components for streaming agent responses and representing work inside channels.

That strategy addresses a basic weakness in standalone coding agents. Software development decisions are scattered across bug reports, product discussions, repository issues, pull-request reviews and deployment logs. The person assigning a task often has to reconstruct that context before the agent can begin. A Slack agent can start from the discussion where the task originated and keep the human reviewers in the loop as it proceeds.

Shared context creates a larger permission problem. Slack's own agent design guidance says agents must respect the visibility boundaries of public channels, private channels and direct messages. Coding agents also need access to repositories, credentials and execution environments, so administrators must decide which conversations can trigger work and which actions require approval.

The commercial logic is straightforward. Anthropic, GitHub, Cognition and Vercel continue selling their respective agent, repository and deployment products. Slack becomes the common front end through which engineering teams use them. Salesforce gains activity and institutional context inside Slack, while avoiding the need to build every coding agent itself.

Slack Code therefore represents an aggregation bet. Coding agents are multiplying, and most engineering organizations will use several. Benioff is betting that teams will manage that collection from the channel where the work was requested, discussed and reviewed in the first place.