Salesforce shipped Slack Code on August 20. Coding agents now run inside dedicated Slack channels, with tabs for the conversation, the agent's proposed plan, a line by line diff, and a live preview of the running output. Claude Code, Devin, GitHub Copilot, ChatGPT and Vercel agents all plug in. A human approves before anything ships, and the whole thing is logged.
The easy read is that Salesforce built a chat interface for coding agents. That's the least interesting part of it.
What actually shipped is a supervision layer. Approval gates and an audit trail, on by default. Most teams running agents right now have neither, and they have not noticed because the agents are still mostly getting it right.
The number that makes this urgent
GitKraken surveyed 554 developers and engineering leaders for its 2026 report. 96.4% of organisations use AI coding tools. The share of developers whose primary way of working is autonomous AI went from 7.6% to 28% in nine months. A third of them keep agents running for the entire workday.
So ask a plain question about your own team. An agent opened a PR last Tuesday. Who prompted it? What plan did it propose before it started? Who approved the diff, and how much of it did they actually read?
Most teams cannot answer any of that. The agent ran on someone's laptop, in a terminal, against a personal access token, and the only artifact left behind is a commit with a human's name on it. That was fine when agents produced autocomplete suggestions. It stops being fine when a quarter of your engineers have something running unattended for eight hours.
You don't need Slack to fix this
The pattern is worth copying even if you never touch the product. Four things, all of which you can do this sprint:
Give agents their own identity. An agent should not commit as a person or authenticate with a human's token. Separate bot accounts, separate scoped credentials, separate branch protection rules. When something goes wrong you want to know it was the agent, not spend an afternoon proving it.
Turn the plan into an artifact. Before an agent edits files, make it write down what it intends to do and where. Post that somewhere durable. A plan you can read in fifteen seconds is the cheapest review you will ever run, and it catches the wrong-approach failures that a diff review catches far too late.
Log the approval, not just the merge. Record who signed off and what they were shown. Not for compliance theatre. For the day a subtle regression ships and you need to reconstruct the decision instead of guessing at it.
Set a blast radius per agent. Which repos, which directories, which environments, which secrets. Default to read-only and widen deliberately. Almost nobody does this, and it's the single control that turns a bad agent run from an incident into an annoyance.
The failure mode nobody plans for
The risk here isn't an agent going rogue. It's much duller than that. It's five people on a team each running their own agent, against their own prompts, with their own context, all merging into the same codebase, and nobody able to reconstruct why the architecture drifted over six weeks.
Slack's bet is that supervision belongs where the team already talks. You might disagree with the venue. The underlying point holds regardless: agent oversight is infrastructure now, and if you don't build it, you are relying on the goodwill of whoever last hit approve.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!