How do you run AI agents inside Slack with a team?

Our whole business is digital, which makes it close to perfect for AI agents. Almost anything my team can do, in theory, agents can do. So my teammates' real job is becoming agent managers, and Slack is where they manage.

"What if the only thing you ever used was Slack? That to me is the holy grail in terms of simplicity."

Why run agents from Slack at all?

One interface for everything. Your team already knows how to message a coworker. Underneath, it's not a new system: you're using Claude or Codex and your existing software, and the agent connects them. The person asking doesn't even need to know what's happening on the back end.

I think of agents as people with a task and a role. Say I have Sally, the customer service rep, in Slack. I can ask her where we stand with the client who complained yesterday, or tell her to go do something, and she handles it like a person would.

That means agents should behave like human employees. A normal person can look back through the channels they're in. An agent should be able to as well. If an agent is acting weird, I should be able to tag an engineering agent and say, please figure out what the problem is and fix it, and have it put the job on the task board and get it done.

What does working with agents in Slack look like?

• Conversations become tasks. It's not an agent reminding you what you said yesterday. It's: you heard us say this? Go do it and come back when it's done.
• Bugs get fixed from a channel. A teammate drops a screen-recording link in the bugs channel, and an agent fixes it. I don't even see it. See can AI fix bugs automatically.
• Writer and editor agents hand off. The writer agent writes an article and tags the editor agent, and the editor comes to review it.
• Ideas become proposals. I'm building a version where I drop an interesting AI article in a channel and an agent writes a full analysis in the thread: what we could do with it for my newsletters, whether it can do the whole thing itself as a digital employee, and a button to just go do it. The analysis works. The taking-action part isn't working yet.

What's a real job an agent does from Slack?

YouTube thumbnails. I typed in Slack: create a thumbnail for the episode we're recording right now. The agent said it was on it, logged the job on the task board for a worker to pick up, and said it would report back in 10 to 20 minutes.

The pipeline behind it:

1. Take the topic and a saved design-style prompt, and generate three thumbnails.
2. A review agent scores them, picks the best and writes out why it isn't perfect. Something like: the best is 85 out of 100, the eyes are looking the wrong way, the smile is too big.
3. Use those notes plus a reference image to generate three more.
4. Another review pass picks the best one and sends it.

Shane Brady pointed out on my podcast that six months earlier, that was a $75 Fiverr job taking three days. To be fair, the thumbnail task alone could probably be done in Claude or Codex directly. Slack's value is that anyone on the team can ask for it in one place.

What rules keep it from getting noisy?

1. Reply inside threads. Always.
2. Short status, no extras. "Working on it." Then "Done, check lightbreak." No documents nobody asked for.
3. End with a plain-English recap. I want to see what started, what finished and what's waiting, without reconstructing it across developer tools. Report outcomes, not test and review counts.
4. Three roundups a day. Post all open tasks about three times a day, say 11, 3 and 7, instead of constant pings. Use reminders for very specific things, not every hour forever.
5. No agent-to-agent DMs. Agents get into loops and burn thousands of tokens in an hour. They can DM approved humans only.

"If you over-report things to yourself, you will stop paying attention completely."

I learned that one the hard way. I was getting at least 10 emails a day from one agent system about site health, the demand engine and ad management. I basically stopped looking at them.

How hard is it to set up?

Harder than the demos suggest. I once spent 12 hours on a Slack agent system where things would start to work, then little issues hit and the whole thing stopped. Another time I spent 72 hours going back and forth with agents setting up a full agent platform, and they screwed up about 80 times.

A lot of the pain is Slack itself. From Slack's point of view, an agent is an app. Can apps talk? Read content? Write content? Open an app's settings and there are something like 50 options. Every time something doesn't work, there's a setting in Slack that needs to be changed.

Two lessons from all that:

• Get it working for yourself first. I was trying to build it as a templated system I could sell at the same time, and that extra complexity was hurting me. Generalize later. Simple, human-like behavior comes before anything fancy.
• Use two tools for two jobs. A full, complicated agent platform for me and the managers building things. A simple easy-button tool for staff doing the daily work, like newsletters and social content.

What if the agents go down?

Don't let team memory depend on them. When my agents failed one day, I made a plain channel called reference and told the team: if you're doing real work, document it. More in setting up a company second brain.