About 10 months ago we moved the whole team at StealthX to Claude Code and went all in on it. It's how everybody here works now, and watching this team get good at it as fast as they did is one of the things I'm proudest of.
We're making the next move, and it's a different shape than the last one. That one was about each of us getting sharper on our own. This one is about the team sharing AI teammates.
If you're using Claude Code, or handing Claude a transcript and getting a document back, that's status quo in September 2026. A year ago it was a real edge. The shift now is handing whole jobs to named teammates who happen not to be human.
1. Two ladders, five rungs each
People talk about single player and multiplayer AI like it’s one switch that you flip. The way I see it there are two ladders.. and you can be near the top of one and on the floor of the other.
Single player is you + AI. Here are the levels:
Chat. You ask, it answers, you paste the answer somewhere else.
Context. You've given it your real work, so the answers stop being generic.
Skills. The things you do over and over are written down once, so you stop re-explaining yourself.
Tools. It can go do the thing now.. pull the report, draft the email, update the record.
Loops. It runs on its own schedule and brings you things you didn't ask for.
Multiplayer is your team + AI. Here are the levels:
Shared skills. Somebody improves one and everybody gets it.
Shared context. One governed body of context the whole team works from, instead of a private version per person.
Shared workflows. The same routines and guardrails for everyone, running without a person kicking them off.
Role-based access. A shared AI teammate (aka a shared agent) that works the same way for everyone but only reaches what your role is allowed to see.
Swarms. Agents that spin up other agents and split the work between them.
There are lots of folks that are a 5 on the 1st ladder (i.e., they’ve mostly figured out single player AI and have built something great only they use) and a 1 on the 2nd ladder (i.e., they don’t have any shared workflows or agents with other folks on their team). Almost every company has someone who’s a 5 on the 1st ladder, and they're usually the person everybody points to as proof the company is ahead.
So when somebody asks for multiplayer, the useful question back is.. which level? Most of the time they're asking for level 2 and picturing 4..
2. Most teams are building a faster horse
Most teams answer this by getting more organized. They map the process, they write the skill, and then they wire the workflow. At this step you use this skill, at that step you apply that workflow and things get more and more elaborate every week.
BUT that's a faster horse.. you're optimizing what you’ve got instead of asking whether anybody even needs to do the task/job at all.
If you’re at level 4 or 5 of multiplayer mode in AI, you stop worrying about the process because an AI teammate is actually owning the outcome with clear role/responsibilities + clear goals and guardrails. You don't tell it what to do or what tool to use at each step. You give it a goal or problem, and it just does it for you.
Side note.. standing up AI-enabled skills and workflows aren't wasted effort. You still need to progress through the levels of the ladder because those pieces are the raw material an AI teammate is made of.
The mistake is stopping at multiplayer level 3 and thinking you’ve got AI figured out.
3. What an AI teammate actually is
Say you need a deck for a meeting. Today somebody opens Powerpoint, thinks about structure, hunts down the template, and burns two hours. Some people might have Copilot/Claude/GPT assist with building out the content and structure, but usually it’s not quite right and they still have to polish.
Now picture Dom. Dom is an AI teammate that builds on-brand decks in your company voice. You @ mention him in Slack (or Teams for Microsoft folks out there).. “Dom, make a deck from the planning discussion we just had to present tomorrow morning.” Five minutes later he’s back with a draft. You give him a couple quick pieces of feedback and he knocks it out. Every deck Dom makes is the same template/structure/brand for everyone on your team because Dom works from the same context every time no matter who asked.
Nobody has to police a template anymore because there's only one Dom. Also, you don’t have variance between decks because people are using different deck-making skills in Claude or GPT.
Or imagine Rio, an AI teammate that is specifically trained on recaps in your company-approved tone. Rio picks up the meeting you just walked out of, drafts the recap and leaves it sitting as a draft for you to send. Then shoots over a Slack message to whoever needs to know, schedules the follow-up email, logs/assigns any tasks that came out of the call in your project management system, all fired off without anybody assembling them by hand.
James is my AI Chief of Staff and I’ve been using him for months. I also have Warden, our code reviewer. Dom and Rio aren't built yet but they're the two I'm standing up now.
4. The agents work with each other
I can tag Warden on any commit or pull request in Github to get a code review, but I never do it.. James does. James tags Warden, Warden comments on what's wrong, James fixes it, then James asks Warden to check it again. That loop closes with no human anywhere in it, and I find out it happened when they notify me it’s ready for final review/approval.
Last week I had a 30-minute whiteboard session about hosting a school group at the CLT Startup House, then walked straight into another meeting. While I was sitting in that one, I spent about 60 seconds telling James to go read the transcript, make a plan, and do what I mentioned in the meeting. 20 minutes later he'd built a multi-page deck, a leave-behind for the kids, a draft email summarizing the plan to send to Chris (Exec Director at the house), and the plumbing for a game that I’d build live with them the next morning.
I delegate out loud now, inside the meeting recording. I'll say "James is going to handle this part" while we're still talking, and afterward he reads the transcript to figure out what I assigned him. I don’t have to ask him or write up a bunch of details. The ask and the context is already there, because I said it in the room.
5. Give it a machine, then decide who can talk to it
An AI teammate lives on a machine (for non-technical folks, this is usually a server computer, a computer you have, or a virtual sandbox), and whatever that machine can reach is what that AI teammate can do. That collapses the permissions problem people try to solve with an enormous matrix into two much easier questions: what does this thing have access to, and who's allowed to talk to it?
Just like when you’re hiring a person.
For example, not everyone can just walk into the CFO's office and look at the books. Same is true for an AI teammate… the CFO agent/AI teammate shouldn’t be available/accessible for everybody. If you give someone access to it, you’re saying that they have clearance for that type of work.
Hopefully that makes sense 😅
Wrapping up
I want everybody at StealthX to have their own version of James. I also want the company running like 30 AI teammates that handle specific jobs every day so they can spend their hours on client relationships/building/thinking.
Currently we have three that work and a list of the rest. If you're reading this from level 1 with no shared anything, that's a completely normal place to be standing right now. Start by picking one recurring task/job to delegate to AI.
Putting this into action
Pick the one job your team redoes from scratch every week and write down what "done" looks like for it. That description is the starting point for an AI agent/teammate, and you can write it today without touching a tool.
If you've got a meeting notetaker tool, try delegating out loud in your next recorded meeting. Say what you want handled, then point your assistant at the transcript afterward instead of writing a brief.
What I'm reading
The Multiplayer AI Sprint, on The AI Daily Brief. Labor Day episode on shared agents and launched a free four-week sprint to get a team to its first one. It's the most usable on-ramp I've seen for the ladder above, so go do it.
Marcus. An AI employee you reach from Slack, Teams, email or SMS. Worth a look as the bought version of what I described building, and worth noticing how much of the pitch is about where you talk to it rather than what it knows.
Onward & upward 🤘
Drew