Your Best Engineers Are Doing AI Work Nobody Can See.

Slack shipped a product on August 20 that says something uncomfortable about how your engineering team actually works.

Slack Code puts AI coding agents into shared channels. Claude, Devin, GitHub Copilot, ChatGPT, and Vercel's agents all run where the team can see them, with side-by-side diffs, live previews, and the whole planning conversation in one place. It works on any Slack plan.

Rob Seaman, the EVP who runs Slack, framed the reasoning in one line: AI only creates value when it is part of how a team actually works, not something people go do alone in another tab.

Read that again with your own company in mind.

The Tab Is the Point

For two years, the most consequential work happening at your company has been happening in private.

Somebody on your team is three hours into a session with a coding agent right now. They are making architectural decisions, discovering which prompts work and which ones waste forty minutes, learning where the model is confidently wrong. That knowledge is real and it is expensive to acquire.

It lives in their terminal. When they close the laptop, it is gone.

This is not a discipline problem. Nobody is hiding anything. The tooling was built for a single operator at a single keyboard, and it made privacy the default. Everything downstream of that default follows automatically.

Nobody Learns Twice

Here is what invisible work costs you, and it is not what most leaders assume.

The cost is not that people are goofing off. The cost is that every person on your team is paying the same tuition separately.

Your senior engineer figured out last month that the agent needs the schema pasted in or it invents column names. Your new hire will discover that next week, on their own time, after producing two broken pull requests. Nobody wrote it down because there was no natural place to write it down, and no one saw it happen.

Multiply that by every person and every workflow. That is your actual AI adoption cost, and it does not show up in any budget line.

I wrote two days ago that judgment is the scarce input now and that execution got cheap. Here is the corollary I left out. Judgment that lives in one person's head is not an organizational asset. It is a single point of failure with a salary.

Visibility Is Not Governance

I want to be careful not to oversell this, because a Slack channel is not a control.

Moving agent work into a shared channel makes it observable. It does not make it safe, reviewed, or correct. You can watch an agent produce bad code in real time with the whole team present and still ship it. You can pipe every session into a channel nobody reads and call that transparency.

Same lesson I have been repeating all month in a different costume. Yesterday it was that a spending limit only counts when something outside the agent enforces it. A channel is not enforcement. It is lighting.

But lighting matters. You cannot review what you cannot see, you cannot standardize what you never observed, and you cannot coach somebody through a workflow that happens entirely off-camera. Observability is the precondition for every control you would want to build later.

Start with the lights. Do not confuse them with the lock.

The Pillar This Sits On

No lone wolves. That is the first of our four pillars and it is the one that took the biggest hit these last two years.

AI tooling has quietly made individual contributors more powerful and teams less connected. The best engineer on your staff can now do the work of three people without talking to anybody, and that feels like a win until you need somebody else to maintain it, or until that person leaves and takes the entire operating manual with them.

Elite delivery was never about one person going fast. It was about the unit shipping reliably, repeatedly, with the knowledge held in common. Systems, not heroics.

An agent working in a private terminal is heroics with better tooling.

What I'd Do This Week

Ask your team where their AI work happens. Not whether they use it. Where. If every answer is a local terminal or a personal chat window, you have found your gap.

Create one shared place for it. Slack Code if you are on Slack, a channel with pasted transcripts if you are not. The specific tool matters far less than the default changing from private to visible.

Write down the three things that keep going wrong. Every team has them. The context the agent always needs, the task it always fumbles, the check that always has to happen after. That is a document, and it should take an afternoon.

Make one person's session public every week. Not as surveillance. As teaching. The fastest way to spread a workflow is to let people watch somebody good use it.

Stop measuring AI adoption by license count. Seats tell you what you bought. Shared artifacts tell you what you learned.

The Part That Matters

Your team already has the reps. They are just scattered across a dozen machines where nobody can learn from them.

The tooling made that the default. It does not have to stay the default.

Turn the lights on. Then decide what needs a lock.