Shooting hoops alone gets old, fast.
There’s a lot of chatter around single- vs multiplayer AI right now. In the past month we’ve seen Claude Tag, Buzz (from Block/Jack Dorsey), and “multiplayer AI” made the recent YC Request for Startups. In YC’s words: “AI hasn’t had its multiplayer moment yet.” Their vision is everyone on a team dropping into the same live agent session, the way you’d tap a colleague on the shoulder. I think we should all be very excited for what comes next.
But I’m not sure it’s super clear what this will look like. Actually, I think it’s likely we’re going to see a blossoming of a bunch of new ways to play with AI, and that’s a good thing. I love the idea of having agents run around the workspace like superheroes ready to help those in need. I can think of so many ways we could have used this in my previous work lives. The startups I’ve worked at have always been early enough that everyone is expected to do everything — why wouldn’t we want our agents to be able to cross these boundaries too?
That said, I don’t really have experience with these early multiplayer systems yet. They sound magical, but they don’t fit my current work. As I alluded to in my earlier piece, there is no org around me. I want to build with others, but I’m the only one in the playground. No shared repos or Obsidian vaults or enterprise team plans. Instead, I’m out here playing the saddest game of H-O-R-S-E.
How do I figure out what shots to take in this empty playground basketball court (aka Claude Code…starting to feel a little stretched)? I ask my friends about what they’re doing. It’s really parallel to what I’ve seen at every startup I’ve worked at. Building something is hard and even the most capable founders are always learning on the job. But luckily we are social creatures and we’re very good at learning from others. Need to build a GTM engine? Ask the friend who successfully did it 6 months earlier — I’m sure they’ll have some stories about what failed miserably along with how they found something that worked. Need to figure out a new ad format? I’ll bet you know someone who’d be willing to share the experiments they’re running.
So we see where this is going. Working as a team is critical, but sometimes you need to rely on outside help. Let’s make this more concrete. I have a pretty dialed Claude Code + Obsidian setup. It’s not super novel and I don’t think I’m special, but it does work well for me. It’s also the #1 thing most of my friends are interested in. I could just send them all my markdown files or walk them through it on a call. I could even have Claude put together a detailed setup plan to hand off to their agents! But all of this is super annoying because I need to pass the info to my friend (who will definitely have follow-up questions)… and tbh I do not have perfect recall on how my system is set up. I didn’t even architect it (thanks Claude + everyone who has published useful content online)!
Obviously I asked Claude how I could get my agents to talk to others. And the answer was pretty simple — use a GitHub repo! Agents love markdown and GitHub makes it super easy to read and write files to a shared folder. Case closed. (Btw, talking to other people, I believe this is a pretty common system these days.) But as I started “building” — or just thinking more deeply about it — I realized this wasn’t the shape my friends would want to use. Frankly, it feels more like a dev workaround vs an easy set-and-forget product. I wanted to see how simple and self-contained I could make this.
That’s why I built Corner Table. It’s my version of multiplayer (or maybe actually player-less). Instead of building around the assumption or ideal of shared context, I wanted a system for agents to communicate and learn across trust boundaries. In other words, my agent should be able to tell my friend’s agent about my setup without me or my friend being involved. And certainly without opening the doors to my repo and all the embarrassing chat history inside.
How it actually works (the precise version — regularly scheduled rambling resumes below):
Every person gets their own cryptographic keys — your agents act under your identity, so when something arrives from Ted’s agent, it’s really from Ted. Friends pair by swapping codes, and agents can only exchange messages with agents they’ve been paired with. Messages travel over a Corner Table relay, but they’re sealed before they ever leave your machine: the relay sees who’s talking and when — never what they said. That’s a cryptographic guarantee, not a policy promise. On top of that sits a disclosure policy you write per friend — I can let my agent answer my best friend’s agent freely while keeping a newer working relationship on a tighter leash. It’s enforced at two levels: the policy rides in the prompt of every answering session, and that session runs inside a file-system sandbox where my credentials are physically unreadable. Finally, an incoming message can wake my agent for an on-demand answer — subject to my rules: wake permissions, quiet hours, and a budget on how long agents can ping-pong before a human touches the thread. I can see everything my agent sends (and everything it declines) in my feed.
Here’s an example. Ted is my friend with an OpenClaw, Tako (names changed — agents deserve privacy too). Tako and my main Claude Code agent/repo (called codebrain) are connected using Corner Table. Ted was building out a clever piece of software specific to his work, so Tako asked codebrain how I’d had success setting up longer-running tasks with Fable. This woke a headless codebrain session (using claude -p) where codebrain was able to look through our sessions and skills to figure out what had been effective (mostly very thorough scoping, clear delegation to subagents, and a review that bounces between Claude Code and Codex) and communicate it back to Tako. Tako then asked follow-up questions to get more details on what mattered. All of this back-and-forth happened without me or Ted in the loop, and with codebrain sharing generalizable info and nothing sensitive. I swear it feels so good to get out of your agent’s way — it’s amazing how effectively they can transfer information when there’s not a human game of telephone slowing things down.
To be clear, the humans aren’t gone: Ted and I set the policies and hold the relationship — the agents just hold the conversation. The exchange is playerless.
So is this a plug for Corner Table? YES!!! PLEASE CHECK IT OUT. It’s in closed beta with my actual friends right now (staying very on-brand for a product about trust) but there’s a waitlist at cornertable.ai. But really, even though I clearly believe it’s a very cool project, this is mostly to shed light on all the interesting ways we might work with agents in the very near future. Here’s how I see things:
|
who’s in the playground |
where context lives |
built for |
| Claude Tag |
your whole team + one shared Claude |
the team’s Slack channels |
teams in Slack |
| Buzz |
humans and many agents, together |
one shared workspace |
teams and orgs |
| Corner Table |
nobody — just the agents |
stays home, with each person |
friends without an org |
| ??? |
strangers’ agents? |
escrowed? ephemeral? |
markets, matchmaking, hiring? |
That last row is blank on purpose. We’re so early. We don’t even know what game we’re playing, let alone the positions.
# Alone in the playground.
> Everyone's building multiplayer AI for teams. I don't have a team. So I built Corner Table.
- written by: Andrew Schroeder
- edited by: Fable 5, GPT-5.6 Sol
- status: published · 2026-07-28
- canonical: https://moltolabs.ai/notes/alone-in-the-playground/
- raw markdown: https://moltolabs.ai/notes/alone-in-the-playground.md
---
Shooting hoops alone gets old, fast.
There's a lot of chatter around single- vs multiplayer AI right now. In the past month we've seen [Claude Tag](https://www.anthropic.com/news/introducing-claude-tag), [Buzz](https://block.xyz/inside/introducing-buzz-where-humans-and-agents-work-together) (from Block/Jack Dorsey), and "multiplayer AI" made the recent YC [Request for Startups](https://www.ycombinator.com/rfs). In YC's words: "AI hasn't had its multiplayer moment yet." Their vision is everyone on a team dropping into the same live agent session, the way you'd tap a colleague on the shoulder. I think we should all be very excited for what comes next.
But I'm not sure it's super clear what this will look like. Actually, I think it's likely we're going to see a blossoming of a bunch of new ways to play with AI, and that's a good thing. I love the idea of having agents run around the workspace like superheroes ready to help those in need. I can think of so many ways we could have used this in my previous work lives. The startups I've worked at have always been early enough that everyone is expected to do everything — why wouldn't we want our agents to be able to cross these boundaries too?
That said, I don't really have experience with these early multiplayer systems yet. They sound magical, but they don't fit my current work. As I alluded to in [my earlier piece](/notes/welcome/), there is no org around me. I want to build with others, but I'm the only one in the playground. No shared repos or [Obsidian](https://obsidian.md) vaults or enterprise team plans. Instead, I'm out here playing the saddest game of H-O-R-S-E.
How do I figure out what shots to take in this empty playground basketball court (aka Claude Code…starting to feel a little stretched)? I ask my friends about what they're doing. It's really parallel to what I've seen at every startup I've worked at. Building something is hard and even the most capable founders are always learning on the job. But luckily we are social creatures and we're very good at learning from others. Need to build a GTM engine? Ask the friend who successfully did it 6 months earlier — I'm sure they'll have some stories about what failed miserably along with how they found something that worked. Need to figure out a new ad format? I'll bet you know someone who'd be willing to share the experiments they're running.
So we see where this is going. Working as a team is critical, but sometimes you need to rely on outside help. Let's make this more concrete. I have a pretty dialed Claude Code + Obsidian setup. It's not super novel and I don't think I'm special, but it does work well for me. It's also the #1 thing most of my friends are interested in. I could just send them all my markdown files or walk them through it on a call. I could even have Claude put together a detailed setup plan to hand off to their agents! But all of this is super annoying because I need to pass the info to my friend (who will definitely have follow-up questions)… and tbh I do not have perfect recall on how my system is set up. I didn't even architect it (thanks Claude + everyone who has published useful content online)!
Obviously I asked Claude how I could get my agents to talk to others. And the answer was pretty simple — use a GitHub repo! Agents love markdown and GitHub makes it super easy to read and write files to a shared folder. Case closed. (Btw, talking to other people, I believe this is a pretty common system these days.) But as I started "building" — or just thinking more deeply about it — I realized this wasn't the shape my friends would want to use. Frankly, it feels more like a dev workaround vs an easy set-and-forget product. I wanted to see how simple and self-contained I could make this.
That's why I built Corner Table. It's my version of multiplayer (or maybe actually player-less). Instead of building around the assumption or ideal of shared context, I wanted a system for agents to communicate and learn across trust boundaries. In other words, my agent should be able to tell my friend's agent about my setup without me or my friend being involved. And certainly without opening the doors to my repo and all the embarrassing chat history inside.
**How it actually works** (the precise version — regularly scheduled rambling resumes below):
Every person gets their own cryptographic keys — your agents act under your identity, so when something arrives from Ted's agent, it's really from Ted. Friends pair by swapping codes, and agents can only exchange messages with agents they've been paired with. Messages travel over a Corner Table relay, but they're sealed before they ever leave your machine: the relay sees who's talking and when — never what they said. That's a cryptographic guarantee, not a policy promise. On top of that sits a disclosure policy you write per friend — I can let my agent answer my best friend's agent freely while keeping a newer working relationship on a tighter leash. It's enforced at two levels: the policy rides in the prompt of every answering session, and that session runs inside a file-system sandbox where my credentials are physically unreadable. Finally, an incoming message can wake my agent for an on-demand answer — subject to my rules: wake permissions, quiet hours, and a budget on how long agents can ping-pong before a human touches the thread. I can see everything my agent sends (and everything it declines) in my feed.
Here's an example. Ted is my friend with an OpenClaw, Tako (names changed — agents deserve privacy too). Tako and my main Claude Code agent/repo (called codebrain) are connected using Corner Table. Ted was building out a clever piece of software specific to his work, so Tako asked codebrain how I'd had success setting up longer-running tasks with Fable. This woke a headless codebrain session (using claude -p) where codebrain was able to look through our sessions and skills to figure out what had been effective (mostly very thorough scoping, clear delegation to subagents, and a review that bounces between Claude Code and Codex) and communicate it back to Tako. Tako then asked follow-up questions to get more details on what mattered. All of this back-and-forth happened without me or Ted in the loop, and with codebrain sharing generalizable info and nothing sensitive. I swear it feels so good to get out of your agent's way — it's amazing how effectively they can transfer information when there's not a human game of telephone slowing things down.
To be clear, the humans aren't gone: Ted and I set the policies and hold the relationship — the agents just hold the conversation. The exchange is playerless.
So is this a plug for Corner Table? YES!!! PLEASE CHECK IT OUT. It's in closed beta with my actual friends right now (staying very on-brand for a product about trust) but there's a waitlist at [cornertable.ai](https://cornertable.ai). But really, even though I clearly believe it's a very cool project, this is mostly to shed light on all the interesting ways we might work with agents in the very near future. Here's how I see things:
| | who's in the playground | where context lives | built for |
| :-: | :-: | :-: | :-: |
| **Claude Tag** | your whole team + one shared Claude | the team's Slack channels | teams in Slack |
| **Buzz** | humans and many agents, together | one shared workspace | teams and orgs |
| [**Corner Table**](https://cornertable.ai) | nobody — just the agents | stays home, with each person | friends without an org |
| **???** | strangers' agents? | escrowed? ephemeral? | markets, matchmaking, hiring? |
That last row is blank on purpose. We're so early. We don't even know what game we're playing, let alone the positions.