Slack has launched a collaborative environment for AI coding agents that moves software work out of private developer sessions and into shared channels, where colleagues can inspect changes, supply context, interrupt an agent, and approve its work before code reaches production.

Slack Code creates dedicated project channels when a user summons a supported coding agent, including products from Anthropic, Cognition, GitHub, OpenAI, and Vercel. The resulting workspace contains the conversation alongside code changes, previews, plans, and other outputs rather than reducing the interaction to a private chatbot session.

When work is complete, the channel can archive while retaining its history as an organisational record. Slack says the feature is available across its plans, although customers still require access to whichever third-party coding agent they choose.

Coding assistants have already altered how individual software developers work. The next problem for enterprise adoption is organisational: an agent operating inside one person’s terminal or browser may produce code quickly, while its reasoning, context, changes, and mistakes can remain invisible until another employee reviews the result.

Agent work becomes visible to the team

Slack Code attempts to make that activity collaborative from the beginning. A product manager can initiate work from an existing conversation, after which an agent creates a project channel, gathers relevant context, proposes changes, and exposes its output to engineers, designers, and other colleagues.

Participants can inspect code differences, run previews, provide feedback, and redirect or stop the agent. Higher-risk actions such as sending code towards production are designed to require human approval rather than allowing autonomous execution without an accountable decision point.

That structure addresses a weakness in individual AI assistants. Traditional copilots are usually optimised around the person using them, whereas organisational processes depend on permissions, handovers, review, shared records, and accountability.

Moving an agent into the collaboration layer allows those controls to surround its work. Slack says agents inherit existing permissions and administrative controls, reducing the need to construct a separate governance environment around every integrated tool.

The proposition is not that chat software replaces development infrastructure. Source control, testing, deployment systems, security tooling, and engineering environments remain necessary. Slack is instead trying to become the place where people coordinate agents operating across those systems.

That is a more valuable position than adding another chatbot to a sidebar. If work begins inside Slack and agents can act across connected tools, the collaboration platform becomes an orchestration layer linking people, organisational context, approvals, and automated execution.

Slack opens the platform to competing agents

The company is deliberately avoiding a single-agent model. Its initial integrations include Anthropic’s Claude, Cognition’s Devin, GitHub Copilot, ChatGPT, and Vercel agents, while Slack says its code-channel APIs will later open more broadly.

That gives Salesforce a position resembling an operating environment rather than a model provider. Competing AI companies can supply the intelligence, while Slack manages identity, context, collaboration, governance, and access to organisational conversations.

Such neutrality can become valuable as businesses resist standardising every workload on one AI supplier. Software teams already use different models and agents for different tasks, while procurement choices can shift as capabilities improve and commercial terms change.

Slack can benefit from that competition if each additional agent gives teams another reason to stay inside its environment. The model supplier owns the intelligence, but the collaboration platform controls the place where people decide what the agent should do and whether its output is accepted.

The same structure can extend beyond software development. Salesforce has indicated that the underlying channel model could later support agents performing other multi-step work, turning coding into the first implementation of a broader collaborative-agent environment.

Enterprise agents need more than autonomy

Much of the agentic AI market has been sold around increasing autonomy: systems capable of planning, operating applications, calling tools, and completing longer tasks with fewer human interventions. Organisational use introduces an opposing requirement because companies also need to know what automated systems are doing.

A useful agent requires enough autonomy to complete meaningful work, while administrators require identity, permissions, logs, approval gates, and a way to terminate the process when its behaviour becomes unacceptable.

Embedding agents into an existing collaboration environment offers one possible compromise. Work can proceed asynchronously while remaining observable, and colleagues who would normally discuss the task can intervene without opening a separate agent-management console.

There are limits. Shared visibility can generate noise if every automated action becomes another message, while giving non-specialists access to coding agents does not remove the need for engineering review. A polished preview can conceal weak architecture, security flaws, maintainability problems, or dependencies that require deeper inspection.

Slack also becomes a more consequential security boundary as agents gain the ability to read organisational context and perform actions in connected systems. Existing permissions provide a starting point, but organisations still have to decide whether an agent should inherit everything a human user can access and how its activity is monitored across integrations.

Those governance questions will become more important as agents spread beyond development. A system changing code, updating a customer record, reviewing a contract, or preparing a campaign needs different permissions and approval thresholds even where the collaboration interface looks similar.

Slack Code therefore represents a larger shift than adding programming functions to workplace messaging. The first wave of enterprise generative AI concentrated on helping one employee create work more quickly; the emerging market is attempting to make autonomous software participate in the processes through which teams create, review, and approve that work.

The commercial prize will go partly to whoever develops the strongest agents, but another valuable layer sits where those agents meet colleagues, receive context, and leave an accountable record. Slack is making a direct bid for that layer.