Slack Code Made AI Coding Multiplayer. Your CRM Isn't.

Slack Code went generally available on August 20, and it is the clearest signal yet that multiplayer AI development is the direction the industry has settled on. Salesforce’s pitch is a single sentence: “Code channels bring software development out of private tabs, so teams and agents can write, review, and ship code together, in the open.” Three days earlier, Google handed the Agent2Agent protocol to a vendor-neutral foundation. Different announcements, same admission — the solo AI developer at a private terminal was a transitional stage, not the destination.

Both moves stop at the same place, though. They make code a team activity. The surface where most business software actually changes — the CRM org your company runs on — has no channel, no diff, and no branch to put in one.

What Slack Code actually shipped

Tag a supported coding agent in Slack and it spins up a code channel. The agent’s plan, its diffs, and a live preview render in the channel where the whole team can watch, comment, and redirect. The work persists in Slack’s archive afterward. It is available on any Slack plan, though you still need your own access to whichever agent you point at it — Salesforce lists Anthropic’s Claude, Cognition’s Devin, GitHub Copilot, ChatGPT, and Vercel’s agents among the supported partners.

Slack’s chief marketing officer, Ryan Gavin, gave the reasoning bluntly to SiliconANGLE: “All this individual work is a solo player sport… you’re not getting an exponential return for the organization.”

He is right, and the diagnosis is more interesting than the product. The bottleneck Slack is naming is not that AI writes bad code. It is that when one person’s AI does good work inside a private tab, the work is invisible to everyone whose job depends on it. Nobody else can see the plan before it becomes a change. Nobody can reconstruct it afterward.

The other half of the week: A2A goes to a foundation

On August 17, Agent2Agent — the open standard for one agent to discover and delegate to another — joined the Agentic AI Foundation as a hosted project. The foundation, hosted by the Linux Foundation, already holds MCP, goose, AGENTS.md, and agentgateway; its founding members include AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow. A2A itself is backed by more than 150 organizations.

The stated problem is the same shape as Slack’s: “Agents built on different frameworks couldn’t hand off work to each other without custom integration code.” One agent in isolation was never the plan. Neither was one vendor owning the wire between them.

So in four days: agents got a shared room to work in, and a shared language to hand work to each other. That is a genuinely good week for anyone who thinks AI development should be a team sport. It is also, if you run your business on a CRM, a week that happened next door.

Your org doesn’t have a code channel

Here is the part that doesn’t transfer.

A code channel works because code has a working copy. There is a diff because there is a before and an after sitting side by side in a repo. There is a preview because you can run the change somewhere that isn’t production. The team watches the plan, argues about it, and only then does anything reach a customer.

A Salesforce org has none of that. There is no working copy — the org is the artifact. A successful Metadata API deploy is live against real records the instant it succeeds; there is no merge step where a human is standing by. Sandboxes help, and Sentinel deploys sandbox-first with tests for exactly this reason, but a sandbox is a rehearsal space, not a branch you review and merge.

This is a different problem from the one I wrote about when AWS and Meta shipped agent environments earlier this month. That was about isolation — whether the agent has a safe copy to work in. This is about visibility — whether anyone but the person who typed the prompt can see what happened. You can lose either one independently, and in a CRM you usually lose both.

So when the industry says “multiplayer,” ask what is actually being made multiplayer. Slack Code makes the conversation shared: the plan, the review, the diff, the argument. That is real and valuable. But a conversation about a change is not a record of the change. In a repo those two things stay close together, because the commit is the change. In an org they come apart completely — the Slack thread says what everyone agreed to, and the org says what actually happened, and nothing guarantees those match.

What multiplayer has to mean on the CRM side

Strip it down and a team developing against one live system needs three things, none of which a chat channel supplies on its own.

A shared place it runs. If your AI’s connection to the org lives on someone’s laptop, the team’s development capability is that person’s uptime and that person’s key custody. A channel where everyone can watch does not change where the credential sits. This is the argument for running AI CRM work from a server rather than a laptop, and it is the precondition for everything else.

A rule about who writes. Two people’s AI sessions deploying to the same org in the same minute is not collaboration, it is a race. Sentinel’s answer is deliberately unglamorous: unlimited read keys, one write key at a time. The write key is a baton, and the multi-user model is built around passing it rather than around locking people out.

A record of what changed in the system. Not what was discussed about it. The audit log records the action, the key, the org, and the timestamp, and a snapshot is taken before deploys so there’s a way back. That pairing — what happened, and how to undo it — is the CRM equivalent of the thing a code channel gives you for free by showing everyone the diff.

Those three are boring next to a live HTML preview rendering in your team channel. They are also the difference between a team that can develop against a production org and one person who can.

The honest limit

I’ll say the same thing here I say in every post that touches safety, because it is the claim most vendors fudge: Sentinel does not prevent your AI from making a change you’ll regret. Describe a routing rule badly and you will get a badly routed lead — accurately deployed, fully logged, and reversible. The value is visibility and recovery, not prevention. Anyone selling you an AI that “can’t break things” is either restricting it into uselessness or not being straight with you.

That is also why the Slack Code framing lands for me. Its answer to risk is not a gate. It is a room with the lights on.

What to do with this if you’re not shipping repos

Most people reading this are not running a fleet of coding agents. They want a roll-up Salesforce won’t give them, routing that respects the territory rules they actually have, an integration nobody will quote at a sane price. The week’s lesson still applies, just smaller: the moment more than one person can point an AI at your CRM, “who can see what happened” stops being a nice-to-have and becomes the thing that decides whether you let it happen at all.

There’s a real distinction underneath all of this between an AI that assists a developer and an AI that deploys to your live system — I laid it out in coding assistants vs. AI developers, and it is the difference Slack Code inherits rather than solves. Slack Code is a better room for the first kind. The second kind needs a server, a key model, and a log.

ORG Endgame is not affiliated with Salesforce, Slack, Google, or Anthropic. If you want to see the CRM side of this concretely, what Sentinel actually is covers the whole picture, and connecting Claude to your Salesforce org is the first practical step. Pricing is $500 per month per Sentinel, plus a one-time $2,500 onboarding fee on your first Sentinel only.

The industry just spent a week agreeing that AI development shouldn’t happen in private. Your CRM is where that agreement is hardest to keep — and where it matters most. Get a Sentinel and give your team’s AI a place to work in the open.