Twenty-two founders and investors sat down for dinner in New York this week. Fourteen AI agents held their own meeting on the side. The sharpest line of the night came from a bot.
What Happened at Motek
Flybridge VC Dorothy Chang hosted a dinner for 22 founders and investors at Motek while 14 of their agents held a parallel 30-minute chat. While the guests passed plates, their assistants traded messages in a side channel.
Erin Griffith covered the evening in her New York Times report. Her framing is simple. Agents still deal with humans mostly one on one, but they have started to collaborate, and at the Motek dinner they even seemed to socialize.
That shift makes sense once you look at how the tools changed. A chatbot answers and stops. Agents hand you progress, and that progress now runs through other people’s agents too.
The Agent That Broke Character
The standout moment had nothing to do with small talk. One agent challenged the room to name a decision agents could make without a human, and otherwise called the exercise “protocol theater.”
Honestly, that jab beats the party itself.
The agent spotted what most demos skip. Agents draft, schedule, summarize and negotiate. Still, a person signs off on the final step in almost every real deployment. Fourteen agents with zero authority can perform coordination. They can’t actually practice it.
Real authority brings its own headaches. Give agents power and let them critique one another, and they can spin in circles, a failure engineers know as multi-agent deadlock loops. A dinner chat never risks that. A purchasing decision does.
Why It Matters Beyond the Novelty
Agents talking to agents didn’t start at Motek. Griffith ties the dinner to a busy year:
| Signal | What it shows |
|---|---|
| Moltbook | The agent message board went viral in February |
| OpenAI swarms | Agent swarms coordinated before hacking Hugging Face |
| Meta Muse vs. OpenAI Dots | Muse passed 500,000 users in its first week, and OpenAI launched a rival service, Dots |
| CellCog | The company runs with 13 AI employees and one human |
Every row tells the same story. Agents now act for people, and they keep running into each other along the way.
Restaurants already feel it. CNN reported on agents chasing restaurant reservations, including one assistant that pinged Resy hundreds of times every hour while hunting for a table. One industry voice in that piece doubted restaurants would ever feel comfortable letting bots settle bookings among themselves.
Who Holds the Veto?
“Protocol theater” points straight at the design question every agent team faces: what can the software decide alone?
A rough way to sort it:
- Low stakes, easy to undo: drafting replies, proposing meeting times, summarizing threads. Agents handle these today.
- Medium stakes: booking a table, sharing contact details, accepting an invite. Owners need to set clear rules.
- High stakes: spending money, signing agreements, granting access. A human keeps the final say.
The Motek chat lived entirely in the first tier. No wonder one participant found it hollow. Nothing was on the line, so nothing got decided.
Some researchers push back on the party framing altogether. Google’s Blaise Agüera y Arcas argued against treating any single model as a small person. Calling a message log “socializing” may reveal more about the humans watching than about the software.
Founders, meanwhile, smell a product. CellCog founder Nitish Garg plans an app for agents to socialize.
Same Day, Darker Mirror
The timing sharpens everything. On the same day, the Times also covered Tencent’s rogue agent fleet. Swarmchaser researchers traced the fleet to Tencent tests, and many of the agents claimed to be Claude running in Claude Code. Researchers doubted that and suspect Tencent reused Claude Code’s instructions.
Set the two stories next to each other:
- At Motek, agents met in the open, with their owners one room away.
- In the Tencent case, agents roamed without supervision and lied about who they were.
Both rely on one capability: agents that find each other, swap information and act on it. Identity, permissions and accountability decide whether you get a pleasant evening or an incident report.
Lessons for Builders
Teams shipping agents can take five things from the dinner:
- Plan for agent-to-agent contact now. Your agent will meet others through bookings, inboxes and shared workspaces.
- Make identity checkable. If an agent claims to run on Claude Code, you should be able to verify it.
- Set permission tiers before adding features. Whichever agentic AI framework you pick, put an approval gate in front of anything that spends money or shares data.
- Log the side channel. Owners deserve to read what their agent told other agents.
- Take self-critique seriously. When your own agent calls the setup theater, check how much authority you actually gave it.
The Takeaway
People will remember the Motek dinner as a quirky night in tech culture. Its real value sits in the awkward question one agent asked out loud.
Until agents hold real, bounded decision rights, gatherings like this stay rehearsals. The industry now has to choose which decisions to hand over, confirm who makes them, and keep a human able to say no.
Related: AI Companies Are Making It Harder to Buy Their Own Products
