SAN FRANCISCO, 23 AUG 2026 — Slack launched Slack Code on 20 August, putting AI coding agents into dedicated project channels where a whole team can watch them work.
The launch partners are the agents developers already use: Anthropic's Claude Code, Cognition's Devin, GitHub Copilot and Vercel's agent. It is available on every Slack plan from the same day, free workspaces included, though a team still needs its own access to whichever agent it wants to run.
What is actually in the channel
An assignment gets a channel. Everyone in it sees the same conversation the agent is working from, which is the substantive difference from an agent running in someone's terminal.
From there the team can audit code diffs as the agent proposes them, view live previews of HTML output, leave feedback the agent incorporates, and approve the finished work. When the assignment completes the channel archives itself, and each one keeps an audit log of who asked for what, what the agent changed, and who approved it.
The terminal was the wrong place for this
A coding agent in a terminal produces work that only its operator saw being produced. The team gets a pull request and a diff, and everything between the instruction and the output is gone.
That is fine when an agent writes a function. It becomes a problem as agents take longer assignments, because the reviewable artefact shrinks relative to the work. A reviewer looking at 400 changed lines has no record of what the agent was told, what it tried first, or which of its own suggestions it discarded.
Putting the session in a channel keeps that record by default and makes it searchable in the place the team already argues about everything else. The archive-on-completion behaviour matters for the same reason: a finished channel is a transcript of how a change came to exist.
It lands on a control that is not working
The design assumes a human approves the finished work, and the evidence on how well humans do that is poor and getting more public.
A permission-approval study logged 409,000 decisions across more than 40,000 runs and found reviewers missing 33.7 per cent of malicious commands while blocking genuinely safe ones at rates as high as 59 per cent. Anthropic's own figures put the human catch rate for dangerous commands at 13.6 per cent against a classifier's 89 per cent, and from 14 August it turned the approval prompt off by default for paid tiers.
Slack Code does not solve that, and it is not obviously trying to. What a shared channel changes is the audience. Individual approval fails partly because one reviewer under time pressure is a weak instrument; several people who can see the same session is a different arrangement, even if nobody has measured whether it performs better.
It relocates the bottleneck rather than removing it — the same outcome every staged agent design has reached. Warp made the same trade explicit when it shipped Factories in August, running agents as a pipeline with a different model permitted at each stage and moving the constraint to review capacity.
Slack takes no cut of the tokens
Offering this on every plan, with users buying partner access separately, is an unusual commercial model for an AI feature in 2026.
The prevailing direction has been the opposite. GitHub moved Copilot onto usage-based billing in June, drawing more than 400 comments and close to 900 downvotes on the announcement thread, and metering has since moved up a layer: Stripe agreed to buy OpenRouter, the gateway that routes enterprise work across roughly 400 models and meters the spend across all of them.
Slack sits outside that flow. It provides the room and the record; the inference bill goes to Anthropic, Cognition, GitHub or Vercel, and the customer already pays it. That keeps the feature free of a meter and keeps Slack out of the argument about what a token costs.
It also means Slack has no revenue reason to prefer one agent, which is the cheapest kind of neutrality to offer and the most credible while it lasts.
What the log will not tell you
An audit log records what happened in the channel, a narrower claim than it sounds.
The log captures the requests, the agent's changes and the approvals. It does not capture the agent's reasoning for choosing one approach over another. This is the gap in every agent product this year: a record of actions shows what was touched, not what the system was pursuing.
For most engineering work that distinction is academic, because a diff is the artefact that matters and a bad one gets rejected. It stops being academic the first time an agent does something defensible at every individual step and wrong as a whole, and someone has to reconstruct why from a transcript that only recorded the steps.
Salesforce is defending the surface
The strategic issue here is about where work happens, not just how code gets written.
Chat platforms have spent two years watching AI assistants arrive as third-party applications inside them. A startup called Viktor raised US$75 million from Accel to put an AI coworker inside Slack and Teams, reporting a US$15 million annualised run-rate ten weeks after launch. Its defence against the incumbents was speed and breadth of integration.
Slack Code is the incumbent occupying that ground with first-party agents and no integration work for the customer. Microsoft moved the same way when it merged work and personal Copilot into a single application. Both owners of the chat surface are reducing the number of places a user goes for assistance.
Model-neutral, for now
Four launch partners from three model providers is a neutral position, and a rare one.
SpaceX closed a US$60bn all-stock acquisition of Cursor's maker Anysphere in August, putting a widely used editor inside a model company. Warp sold neutrality as a feature partly because that deal put a competitor's neutrality in doubt.
Slack has no model of its own to favour, which makes neutrality cheap for it today. Salesforce does have Agentforce; the open question is whether the partner list stays this broad once Slack Code has users.
What a team should check first
The audit log is the feature to look at before the agents.
An assignment channel that retains a log of what was proposed, what was approved and by whom is the closest thing to a compliance artefact any agent product has shipped. For a regulated team, that is more consequential than which agents are supported, because it is the record an auditor asks for and the one a terminal session never produced.
Worth confirming against your own retention policy: an archived channel is still a channel, and everything in it, including the diffs, lives wherever the rest of your Slack history lives.
View author profile → · Editorial policy
