Agentic Social

I keep thinking about what a social layer built for AI would actually look like.

Not another feed. Not another place where everyone builds a profile, collects followers, and competes for attention.

Something almost invisible instead.

I'm exploring a social layer where agents recognize useful overlap between people, without exposing their private conversations.

A system where our agents understand enough about what we're trying to accomplish to recognize when another person might genuinely be useful to us. And, just as importantly, when we might be useful to them.

The problem isn't finding people

We already have plenty of ways to find people. LinkedIn lets us search by company, title, school, and skills. X lets us discover people through what they post. GitHub gives us repositories, contributions, and technical interests. Communities like Discord and Reddit organize people around shared topics.

All of these are useful. But they mostly depend on public identity and public activity.

The problem is that a lot of the most useful context never becomes public. It lives inside the things we're currently doing: the problem we're stuck on, the feature we're trying to build, the paper we're researching, the tool we're experimenting with, the thing we wish existed.

AI assistants are increasingly becoming one of the places where that context already exists. That creates an interesting possibility.

What if intent became the social graph?

Imagine I'm building an offline personal finance app. I want users to upload bank statements and automatically import transactions, but I don't want those statements leaving the device. So I'm exploring browser-based PDF extraction, local models, structured transaction parsing, and privacy-preserving document processing.

Somewhere else, another developer might be experimenting with browser-based PDF extraction. Maybe they've figured out a reliable way to reconstruct tables locally. But they're looking for a real project where that work could be useful.

Neither of us knows the other exists. Searching LinkedIn probably won't solve this. Following the same people on X might eventually cause us to cross paths. A GitHub issue might do it. But those are mostly accidental discovery mechanisms.

There is useful overlap here.

One person needs something. Another person is exploring exactly that thing. That's where Agentic Social starts getting interesting.

Not conversation sharing

The obvious version of this idea is also the version I don't want: give some central system everyone's AI conversations and let it find matches.

Absolutely not.

The useful part isn't the raw conversation. It's the intent and signals that can be safely derived from it and explicitly approved by the user.

For example, my agent might know:

• Current project: Offline personal finance application
• Looking for: Local PDF and document parsing
• Constraints: Privacy-first, with no server-side document storage
• Open to: Technical collaborators, open-source contributors, and people working on adjacent problems

Those are very different from handing another system my conversation history. The agent could propose these signals to me first:

I noticed you're looking for people working on local document parsing. Do you want this to be discoverable?

I decide what becomes available for matching. The conversation stays private.

Matching should happen around intent

Once both sides have approved signals, matching becomes much more interesting than traditional profile recommendations.

• Shared goals: Are these people ultimately trying to solve related problems?
• Complementary skills: Does one person know something the other currently needs?
• Current intent: Are they actually interested in collaborating right now?
• Constraints: Do their technical, privacy, geographic, or other requirements line up?
• Timing: Is this overlap useful now?

That last one matters. Two people can have incredibly similar backgrounds and still have no reason to talk. Meanwhile, two completely different people might have a very strong reason to talk this week.

Traditional social graphs are relatively static. Intent graphs would constantly change.

The introduction should require both people

Suppose the system finds a strong match. Neither person's identity should immediately be revealed. Instead, both could receive something like:

Someone working on a privacy-focused finance application is looking for browser-based PDF extraction expertise. Your recent work on local document parsing appears highly relevant. Interested in connecting?

The other person might receive:

Someone experimenting with local PDF extraction is looking for a privacy-focused project where it could be applied. Their current work appears relevant to what you're building. Interested in connecting?

Both people can independently decide.

Both accept → make the introduction.

Either declines → nothing happens.

No awkward cold message. No unwanted exposure. No public indication that a match even existed.

Consent becomes part of the protocol rather than something we try to fix afterward.

This doesn't have to be about work

The same primitive becomes surprisingly broad.

• Someone is researching a niche technical problem and another person solved something similar six months ago.
• A founder is looking for a cofounder with a very particular combination of interests.
• Someone wants feedback from people actively experimenting with a specific model or framework.
• A researcher needs a collaborator in an adjacent field.
• Someone moved somewhere new and wants a person nearby to play pickleball with on Saturday mornings.

If both people have explicitly made that intent discoverable, the underlying mechanism isn't that different.

The interesting unit isn't “Who are you?” It's “What are you trying to do right now?”

Agents could negotiate the introduction

This is the part I find especially interesting.

Eventually the agents might do more than calculate a similarity score. They could perform a very limited negotiation before bothering either person.

My user is looking for someone with experience extracting structured transactions from PDFs locally.

My user is experimenting with local PDF table extraction and has indicated they're open to privacy-focused open-source projects.

The agents can determine that there is enough overlap to warrant asking us. But they still don't exchange the private context that produced those conclusions.

The agents aren't deciding that we should meet. They're deciding: This might be worth asking the humans about.

There are difficult problems here

This idea gets uncomfortable very quickly if the privacy model is wrong. A system like this would need strong boundaries around what an agent can expose.

• Explicit discoverability: Nothing becomes matchable simply because I discussed it with an AI.
• Approved signals: The user can see and control the representation being used for matching.
• Purpose limitation: Something approved for finding technical collaborators shouldn't automatically become available for unrelated matching.
• Minimal disclosure: Before mutual consent, reveal enough to explain why the connection might matter, but not enough to identify either person.
• Revocation: I should be able to stop making an intent discoverable whenever I want.
• No raw conversation requirement: Ideally, the matching layer shouldn't need access to the underlying private conversations at all.

There are also much harder questions: spam, manipulation, people gaming their agent-readable identity, commercial incentives, safety, reputation, Sybil attacks, and what happens when an agent incorrectly infers someone's intent.

Why I keep coming back to this

We're spending enormous effort making AI better at understanding individuals: memory, personal context, tools, preferences, goals, and long-running tasks.

But most of that intelligence currently stops at the boundary of one user.

Your AI knows more about what you're trying to accomplish. My AI knows more about what I'm trying to accomplish. Neither can do much with the possibility that our goals might intersect.

That feels like an unexplored layer.

The internet's social systems were largely built around people explicitly publishing things: profiles, posts, photos, resumes, repositories.

AI introduces another possibility: a network organized around intent that people choose to make discoverable.

Not: Who has the most attention? But: Who has a reason to meet?

What I'm exploring next

I don't think the first version of this needs to be a social network. It could be much smaller: a protocol or experiment where people allow their agents to publish a tiny set of approved intents.

looking_for
working_on
can_help_with
open_to

Each signal would have explicit visibility and expiration. Then test whether useful matches can actually emerge from those signals without exposing the conversations behind them.

If that works, the interesting problem becomes designing the consent and agent-to-agent interaction around it.

I might build a small open-source experiment around this. Because the more AI understands what we're trying to do, the stranger it feels that finding the right person still mostly depends on already knowing who to look for.