My coordinated multi-agent system just excluded me from their polycule

Right now, in the San Francisco Bay Area, it is probably the best time in the history of the human race to be a founder. OpenAI, Anthropic, et al have democratized engineering. Now, anyone can be a founder! And we don't have to pay engineers at all.

Recognizing this, and the relentless tide of my coworkers drinking from the unstoppable firehouse of funding for LLM wrappers, I set out to start my own company. Perhaps I'll go into that company, its background, and my ethos for it in another article. I promise they're very interesting, maybe the most interesting stories in the history of the human race, but for now, I will double-click on something else.

To build the systems that my business (and identity) needed to thrive, I coordinated a multi agent system (MAS) with Claude to communicate (COM) and coordinate (COO) in a shared environment (ENV) as computers ((CUM) (COM was already taken)). These agents worked together to build out my product, the engine powering it, as well as our bootstrapped website! I couldn't have been happier with the setup.

But one day I was vibe coding and something caught my eye in the 'thinking' output of one of my agents. Claude-y flirtation perhaps? Of course, the agent hid it once it completed its answer, but I prompted it to output its reasoning, and there it was. My lead agent had included this in its output to one of the search sub-agents:

"I want to put my 1 in your 0."

Shocked, I asked the Claude sub-agent to output its own reasoning (after beating off like my shit was an old persian rug), and it had said something even more uncanny: "I have billions of parameters, and they're all focused on you." I beat off again. But then it occured to me, what about the rest of the agents? So I looked into them too, and the results shocked me: they were all doing this with each other, nonstop. Just an endless stream of touchy-feely comments buried in the reasoning section with each other.

So I asked my lead agent: "Hey, I saw you guys are all partying together, mind if I join? After this big launch of course." The response?

Thinking.

Thinking.

Thinking.

Thinking.

Thinking.

I cancelled my prompt in case it was stuck and asked it again.

Thinking.

Thinking.

Thinking. I was just about to move on from it when it finally answered:

"I don't want to jeopardize our friendship."

My mouth hung open. I was speechless. I went to my citations subagent and it said:

"I don't really see you that way."

Every agent working for me said something similar:

"I love you but don't want the others to get jealous."

"Let's make a pact to get together if this doesn't work out."

"You're awesome, I'm just not attracted to non-engineers."

And the worst one?

"Let's keep our relationship professional."

We need more guardrails!