Product4 distinct publishers3 min readUpdated
Slack shipped agent-run code channels on every plan. The bug-to-ship path can now skip the human engineer, and only "high-stakes actions" are gated.
The Product Desk · Product desk

Compiled by The Product DeskSomething wrong?How this is made
Slack has launched Slack Code, a feature that turns software-building AI agents into taggable participants inside team channels: tag an agent and it spins up a dedicated code channel, pulling in the relevant people and context to work the task in the open [1][2][3]. It is available immediately on any Slack plan, with founding-partner agents Claude Code, Devin, Vercel Agent and GitHub Copilot, plus other agents from Slack's marketplace [4][5].
The consequence is in the routing, not the code generation. Gizmodo describes the intended flow plainly: a project manager who hears about a bug from a colleague can tag Claude Code, Devin, GitHub Copilot or a similar agent and have the fix built and shipped without waiting for a human engineer who may be busy with something else [6]. That is a ticket-to-production path in which no engineer is structurally required to appear.
Slack has put a gate on it. Company spokesperson Gianna Dimick told Gizmodo that Slack Code requires human sign-off on high-stakes actions such as merging code to production, governed by Slack's built-in enterprise security model [7]. In its press release, Slack says everyone in the channel can audit code diffs, get live previews of the agent's output, give feedback, and approve the work before it ships [3]. The channel then archives itself when the job is done and leaves an audit log for recordkeeping [10][11].
Read those two statements together and the gap is a matter of who, not whether. The sign-off is described in terms of the team approving [3][10]; nothing in the press release quote or the spokesperson statement, as reported, defines which humans are qualified to give it [13]. In a channel deliberately populated by the project manager, the requester and whoever else got auto-invited, "the team approved" and "someone competent read the diff" are not the same event. Gizmodo's own aside is the reason this matters: agents are prone to misinterpreting instructions and deleting codebases [8].
Slack is explicit that the silo is the thing it is removing. "This isn't your traditional coding experience, with one person and one agent working in a silo on their own," said Katie Steigman, Slack VP of product, in the demo video. "In the Code channel, the whole team and the agent work together on the build" [9]. Marc Benioff, whose Salesforce acquired Slack in 2020, posted the demo on X on August 20, 2026 with the line "Humans and agents. Same channel. Same work" [12][14].
So write the policy before the first channel opens. Name which repositories and environments are eligible; name the roles whose approval counts on a merge and whether a PM's thumbs-up qualifies; decide whether the gating that Slack applies to founding-partner agents also holds for arbitrary marketplace agents [4][7]; and confirm where the audit log lands and how long it is retained, given the channel deletes itself from view on completion [10][11].
What to watch: whether Slack's sign-off requirement is role-aware or merely human-aware, how the diff-and-preview review surfaces behave on changes bigger than a web page [15], and whether teams keep the archived channels discoverable when the first agent-shipped change needs a post-mortem.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Slack debuted a feature called Slack Code, which lets teams collaborate with software-building AI agents directly within group chats.
Users can tag an agent when they need automated assistance; it then starts a new chat, automatically including all the relevant team members and the context it needs to complete the task.
Slack's press release: "With Slack Code, when you have an idea or need to build a new feature, update a web page, or fix a bug, you simply tag in a coding agent like Anthropic's Claude or Cognition's Devin, and that agent then spins up a code channel to tackle the task... There, everyone has full visibility to the conversation, can audit code diffs, get live previews of the agent's output, give feedback, and approve the work before it ships."
Slack Code is available starting today on any Slack plan.
Slack VP of product Katie Steigman said in the demo video: "This isn't your traditional coding experience, with one person and one agent working in a silo on their own... In the Code channel, the whole team and the agent work together on the build. There's no specialized tools needed, and it's all right here in Slack... Nobody's out of the loop, nobody needs an extra meeting. Work just feels more efficient."
Users can follow the agents' progress in the channel as they build, then review an HTML preview when it is ready; once the team approves the finished product, the agent automatically archives the channel.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Launch mechanics well sourced, outcomes untested
Four independent publishers report the same feature set within hours, with direct press-release quotes, a named company spokesperson, on-record executives (Steigman, Seaman, Gavin, Wu, Rodriguez) and an embedded post from Salesforce's CEO. The product mechanics, availability and governance statements are therefore firmly documented. Evidence weakens on two axes: the founding-partner list differs between accounts, and no source offers independent testing, benchmarks or customer verification of the productivity or review-quality claims.
Day-one availability, no usage evidence
Adoption evidence is entirely supply-side: general availability on every Slack plan from launch day, named founding-partner integrations, an existing marketplace of more than 500 AI apps and agents, an Add to Slack onboarding flow, and Anthropic's earlier Claude Tag surface. No source reports a single customer, deployment, seat count, usage figure or pilot result, and partner agent access is a separate purchase that gates real usage.
Capability real, benefit and safety claims run ahead
The shipped mechanics are concrete and consistently reported, so the gap is not about whether the feature exists. It is about the claims layered on top: 'work just feels more efficient', 'accelerated learning', compressed idea-to-deployment cycles and a bug-to-ship path that skips the engineer are all vendor assertions with no measurement, and one source contradicts the engineer-free framing. On the safety side, sign-off is asserted but no approver role or permission model is specified, and TNW's observation that a thumbs up in a busy channel is not a code review points the other way. Gizmodo's own 'assuming it works' hedge captures the distance between claim and demonstration.
Vendor-timed launch with layered commercial interests
The information flow is almost entirely vendor-originated: a Salesforce press release, a CEO promotional post pointing to Dreamforce, and quotes from Slack, Anthropic and GitHub executives who all benefit from the narrative. The commercial structure is explicit — Salesforce gives away the channels while partner agent vendors monetize access, and Salesforce's own reported $300m Anthropic token commitment plus internal Agentforce budget overlap shape which partners get promoted. SiliconANGLE adds an analyst endorsement inside a vendor-interview frame. No independent customer, auditor or security researcher appears in the record.
Facts firm, consequences unproven
High confidence that the feature exists as described, is available on all plans, and that Slack has asserted a human sign-off requirement — four publishers converge on these within a few hours of each other. Confidence drops on the partner list (a four-versus-five discrepancy), on the precise launch day (Wednesday-evening post versus 'launched on Thursday'), and sharply on whether the workflow really removes engineers or preserves review integrity, where sources conflict and no measurement exists.
leadership
Slack Code makes the chat window a coding surface, and a platform call for engineering leaders1 distinct publisher
product
Wu says Cognition is not for sale. The more useful fact is who bought Cursor last week.1 distinct publisher
build
Gartner: agent inference cost rises 5x by 2028, so budget per workflow, not per model1 distinct publisher
build
Grok 4.6 lands in Copilot two days after launch, and the model picker becomes a procurement problem1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
gizmodo.com
1 article · August 20, 2026
siliconangle.com
1 article · August 20, 2026
thenextweb.com
1 article · August 20, 2026
theverge.com
1 article · August 20, 2026