Software development is rarely a one-person job. A request or an idea may begin with a product manager or another employee, before moving through developers, designers, reviewers, and other stakeholders on its way to becoming working code.
Salesforce now wants to put an AI coding agent directly into that workflow to reduce the number of handoffs and additional time involved in moving a software project from an idea to a finished product.
The CRM software provider has released Slack Code, a new service built around dedicated channels inside Slack where a coding agent can be asked to take on a software development task while the stakeholders involved follow its progress, provide context and intervene as needed.
When a user tags a supported coding agent, such as Anthropic’s Claude, Cognition’s Devin, GitHub Copilot, ChatGPT, or a Vercel agent, from any Slack conversation, the new service automatically creates a project-specific code channel for the task, including relevant team members, Katie Steigman, VP of product management at Slack, said in a short video presentation.
The channel can retain relevant context from the conversation and give participants separate views for the agent’s conversation, its plan, code diffs, and the option to review the changes the agent makes, provide feedback and steer, or stop the agent if needed, Steigman added.
Once the task is complete, the agent can package the work for human sign-off and, depending on the workflow and agent, open a pull request as well, Steigman further said, adding that the code channel then archives automatically, while retaining the history as a searchable record.
That proposed workflow, according to Salesforce, is differentiated from the traditional one-person, one-agent approach to AI coding, where an employee or product manager flags an issue or idea in a collaboration tool like Slack, raises a ticket in a project management tool such as Jira, and then hands it over to a developer who works with a coding agent in an IDE, terminal, or another interface.
Salesforce’s bet is that allowing the stakeholders involved in a project to work alongside the same agent and intervene while it carries out a task can reduce the handoffs involved in moving the resulting code through subsequent stages of the development process.
For enterprise teams, Salesforce’s proposed workflow model via Slack Code can offer productivity benefits, analysts say. “The model around Slack Code could mean faster fixes, better cross-functional input, and fewer costly handoffs between technical and non-technical staff,” said Manoj Chandra Jha, principal analyst at Nord-IQ Research.
Primarily because putting all the stakeholders involved in a task alongside the same AI agent could help teams catch incorrect assumptions early, before they surface later in the review process and require work to be redone, which adds costs and time, echoed Amit Kumar Jena, AI development manager at IT consulting firm Kanerika.
The same model could also broaden who gets to initiate software development work, Jena said, allowing, for example, a product manager who spots a bug to tag an agent and watch it produce a first pass instead of filing a ticket and waiting for the task to be picked up.
Beyond the productivity gains, making an agent’s prompts, plans and corrections visible could also help teams learn faster from its work, according to Advait Patel, senior site reliability engineer at Broadcom, particularly by giving other engineers and stakeholders more insight into how the agent was directed rather than limiting their view to the final code diff.
“Stakeholders watching how a good prompt gets constructed and corrected is genuinely useful, and right now none of that is observable,” Patel said.
That same shared context, which is archived as a searchable record after a project is completed, according to Patel, could also come in handy when teams need to revisit or investigate an older code change.
Typically, in traditional workflows, the reasoning behind an agent’s work, including what it was asked to do and the plan it followed, may remain in a developer’s terminal session or across different tools, leaving engineers to reconstruct that context from the final code diff alone.
However, putting an agent’s work in a shared, always-visible channel could also introduce new trade-offs.
“Coding is deep work, and Slack is the interruption machine. Putting them on the same surface is not automatically a win,” Patel said, warning that allowing multiple stakeholders to intervene could also create competing suggestions and midstream changes, potentially slowing a task while forcing the agent to repeatedly adjust to new instructions.
The senior SRE also questioned whether developers would be willing to expose the messier parts of working with an AI agent, including false starts and ineffective prompts: “If developers carry out the messy part privately and paste the clean result into the channel, it defeats the whole purpose. The transparency you get is the performance, not the process.”
There are also concerns around scaling Salesforce’s approach to an extent.
Since Slack Code, according to Patel, creates a new channel for each coding task, that could leave large enterprises managing hundreds of additional channels and create a notification problem that has yet to be addressed.
Those practical concerns and developers’ resistance around making their coding activity visible could also shape how enterprises ultimately adopt Slack Code.
Rather than replacing private developer-agent sessions altogether, Patel expects a hybrid approach in which shared channels are used for work that benefits from cross-functional input, while more focused engineering tasks continue to be handled privately.
“The private loop stays dominant for anything hard. Refactors, debugging, performance work, anything touching a system, you have to hold in your head. Multiplayer fits small, well-scoped, visual or cross-functional work. A copy change, an internal tool, a prototype, a bug, a product manager can describe precisely,” Patel said.
The question of which tasks belong in Slack is only one part of the enterprise adoption equation, however. Security and access controls will also be critical factors, analysts pointed out.
While Jena said enterprises will need to determine what information an agent can access inside a live channel and whether that access is limited to a specific task, Patel cautioned that Slack channel membership should not automatically translate into permissions to access or modify an underlying code repository. Otherwise, adding someone to a channel could effectively become a way to grant code access, creating a potential privilege-escalation risk, Patel noted.
For now, Salesforce said that agents operate within Slack’s existing permissions, security model and administrative controls. The service also comes with APIs that enterprises can use to tailor agent behavior and automate parts of the workflow, including account provisioning and authentication.