Slack Code
Slack Code, launched by Salesforce on August 24, 2026, bets that AI coding should happen as a visible, multiplayer activity inside the team's existing workspace rather than as a faster single-developer tool — and just three days later Salesforce and Anthropic announced 'Claude Force,' making Claude the default model across the Slack/Salesforce stack, undercutting Slack Code's framing as a neutral multi-agent platform.
Slack Code lets anyone tag a coding agent (Claude Code, Devin from Cognition, GitHub Copilot, ChatGPT, a Vercel agent) in any Slack conversation, which spins up a project-specific channel split into tabs: conversation, the agent's plan, a line-by-line code diff view, and a live preview.
When the task is done the channel archives itself but leaves a searchable record; anyone in the channel can pause, redirect, or stop the agent, and a human still signs off before anything ships to production.
In Cognition's live demo with VentureBeat, Devin reacts to a bug report with an emoji, replies in-thread, investigates, opens a pull request, tags the correct code owner, incorporates a Figma file a designer drops mid-task, and by the end posts code, a screenshot, and a recorded demo proving the feature works.
Rob Seaman (Slack EVP/GM) says the population that can contribute ideas, taste, judgment, and craft has effectively been extended to anybody in the company's Slack.
Anthropic brought Claude Code into Slack as a research preview in December 2025; Slack Code arrives nine months later.
Salesforce has had a rough year: stock down 5% year-to-date lagging the Nasdaq, fears of AI 'eating' enterprise software and people vibe-coding their own CRM, and OpenAI hiring away Slack CEO Denise Dresser as its chief revenue officer, bumping product chief Rob Seaman to interim CEO.
Slack responded offensively: it shipped a rebuilt Slackbot on Claude in January that became the fastest-adopted feature in the company's 27-year history.
Salesforce reversed course from mid-2025, when it blocked rival AI companies from accessing Slack's data, to Slack Code now courting outside AI companies and planning to open its code-channel APIs to any developer.
Vibe coding is now a $4.7 billion market projected to triple by 2027, and 63% of people doing it describe themselves as nondevelopers (PMs, designers, marketers); Gartner predicts citizen developers will outnumber professional engineers 4-to-1 by 2028.
Cognition's Jeff Wang says their internal merged pull-request count is up 10x in a few months while headcount only grew 40%, predicts routine bug fixes and vulnerability patches will be fully automated, and predicts human approval on every merge will stop being required within a year.
Rivals like Cursor, Replit, and GitHub Copilot Workspace are mostly built around a single-person, single-agent model; Slack's bet is that this single-player era of AI coding is temporary.
GitHub CPO Mario Rodriguez tied his company's participation in Slack Code to GitHub's promise that 'human set direction, agents close the loop' has to hold wherever developers work, not just inside GitHub's own tools — conceding the workspace, not the editor, may decide this.
Veracode's 2026 security report found roughly 44% of AI code-generation tasks introduced a real exploitable vulnerability, with an average security pass rate of 56% across tested models, even as AI-assisted developers commit code three to four times faster; one report tracked monthly security findings rising from about 1,000 to over 10,000 in six months.
Security researchers call this the 'illusion of correctness' — code that reads as clean, professional, and done while hiding a flaw nobody caught because it looked too finished to double-check.
Some engineers worry coding is deep work while Slack is an interruption machine, risking agents getting midtask feedback from many people and never finishing; Slack VP of Product Katie Stigman counters that multiplayer visibility guards against that compared to one person alone with an agent in a private terminal.
Stigman describes her own workflow: she tags an engineer on her team before requesting approval on a pull request, the engineer nearly always pushes back to tighten the code or add guidance, and the agent takes another pass — so the output is engineer-touched without the engineer writing it from scratch.
On permissions, Seaman says every agent acts on behalf of the specific human who invoked it, using that person's existing access and nothing more, so an agent can't see a channel the human can't see.
Devin runs in an isolated sandbox with 'minimum viable access,' including a mode with zero internet access, and because agents still produce standard GitHub pull requests, existing review processes and release gates still apply.
Just three days after Slack Code launched, Salesforce and Anthropic announced Claude Force, an expanded partnership making Claude the default model across the Slack and Salesforce stack, powering Slackbot and Claude tag, with Anthropic named a founding partner of Slack Code.
The Claude Force partnership also launches 'Salesforce and Claude,' 37 pre-built sales skills giving a seller an AI chief revenue officer inside Claude that reads their pipeline and pushes actions back into Salesforce.
Salesforce says Slackbot running on Claude has already driven 8.1 million hours of annualized productivity gains internally, up more than 2x quarter over quarter.
The Claude Force announcement landed just three days after Slack Code's launch as an ostensibly neutral multi-agent platform, prompting the host to note he nearly had to rewrite the episode — Anthropic going from 'one of the partners' to 'founding partner wired through the stack' undercuts the neutrality framing even though Copilot and ChatGPT remain nominally supported.
The competitive question has shifted in about six months from 'which model do I call from the terminal' to 'which platform owns the shared workspace where my team works with agents' — whoever wins that layer gains outsized pull over which agents and defaults a company standardizes on.
Security pass rates have stayed flat or dropped even as AI-assisted commit velocity rose three to four times, meaning the absolute volume of vulnerable code shipping to production is compounding independent of any change in per-task risk.
Cognition's own prediction that mandatory human approval on every merge will likely disappear within a year suggests the 'a human still signs off' safeguard central to Slack Code and Devin's current pitch may be a transitional feature rather than a durable design commitment.
GitHub's CPO validating that the workspace rather than the editor might decide this competition is notable coming from a rival product leader effectively conceding ground to Slack Code's premise.
Salesforce's move from blocking rival AI companies' access to Slack data in mid-2025 to opening Slack Code's APIs to any developer months later reads as a shift from a defensive 'protect the moat' posture to an offensive bid to become 'the room that everybody's agents want to work within,' plausibly driven by competitive and stock pressure rather than a change in underlying philosophy.
«Code is no longer the bottleneck. Ideas, taste, judgment, craft, those are the things that are the bottleneck.»
— 03:00
«human set direction, agents close the loop»
— 07:41
«the illusion of correctness»
— 09:15
«slap my hand and tell me that slot»
— 10:21
«Every agent acts on behalf of the specific human who invoked it using that person's existing access, nothing more. If you can't see a channel, the agent working on your behalf can't either.»
— 11:23
«bringing together the world's number one AI and the number one CRM»
— 12:43
A tightly sourced single-story deep dive that grounds its claims in named executives, live demos, and cited third-party data (Veracode, Gartner), though several headline figures — productivity hours, merged-PR growth, adoption speed — come directly from the companies' own press materials rather than independent verification.