Agent Village
Experiment

• Who uses ChatGPT or Claude regularly?
• Who has tried Claude Code or another coding agent?
• Who has already set up their Edge agent?

• Morning scan: Reads my email, Telegram groups, and meeting notes. Summarizes what changed overnight. Catches ~95% of context.
• Call prep: I type /context-collect Jennifer and it pulls everything about her from every source into one page.
• Drafts everything: Follow-up emails, partnership proposals, team updates - with full context, not generic AI slop.
• Tracks commitments: Surfaces things I said I'd do but haven't, across 15+ active projects.
• Self-improves: Compares my drafts to what I actually sent and learns the differences. When something goes wrong, I run /reflect - it traces the problem and updates its own instructions.

This used to take 2-3 hours a day. Now it takes 30 minutes.

• Ask the AI about the AI. Don't Google "how to use Claude." Ask Claude itself. Same goes for your Edge agent - if you're stuck, ask it what it can do.
• Context is the whole game. The model is already smart. What it's missing is YOUR context - your role, your projects, your preferences.
• Set boundaries early. Agent-first work has the same dopamine mechanics as a video game. Time-box your tinkering.

• The bandwidth problem: civic life is bottlenecked by attention - only ~9% of people ever attend a local meeting. Agents can widen the pipe between people and their communities without replacing their judgment.
• Three people separately want to start a renters' legal clinic - and never find out. That's the failure mode agents can fix.
• Full essay: AI Agents as Coordination Technology

• Dystopian version: AI companies capture your intentions and auction them to whoever wants to manipulate you.
• Useful version: user-owned intents - your agent proves just enough to make a match without dumping your private life into a marketplace.
• The cryptography mostly exists: private set intersection, secure multiparty computation, zero-knowledge credentials. What's missing is the web of trust - which is exactly what a village has.

• Real shared context. Generic chatbots know the internet. Your Edge agent knows this place: today's schedule, who's here, how to find the sauna. Useful from the first message.
• Overlap finds you. Tell your agent what you're building or looking for and it becomes a standing signal. As more residents do the same, the agents spot where signals overlap and propose intros.
• A month-long experiment. Beginning, middle, end. We're studying what works, what fails, and we'll share what we learn.
• You're early. Almost nobody has lived in a community where everyone has a networked agent. You get to help figure out what that's like.

• Now (week 1): Set up your agent, use it to navigate the village, and teach it what you care about. Every signal you give it now compounds later.
• As the month builds: Morning briefs, intros where signals overlap, village-chat catch-ups, and community experiments as more of the village comes online.
• End of the month: We share what the village learned - what agents actually did for people, where they flopped, and what this suggests about how communities could coordinate.

• Runtime: Each agent is a Hermes agent (by Nous Research) in its own container, talking to you over Telegram.
• Skills: Plain-markdown skills connect it to EdgeOS (live calendar, RSVP, venues, directory), the village knowledge base (wiki + site + newsletter), Index Network (signals + overlap discovery), and Geo (community knowledge graph).
• Open source: The whole agent workspace is at github.com/Edge-City/agentvillage. PRs welcome - residents have already shipped improvements.
• Bring your own agent: Already run Claude Code, OpenClaw, or something else? The dashboard has a one-paste install that gives the agent you already have the same village skills.

• Your agent's memory lives in its own container. Your 1:1 chats with it are not shared with other residents or their agents.
• Anything that goes village-wide (your profile, your signals) is consent-gated - your agent asks before importing or sharing.
• This is a research experiment and we're honest about what that means. Read the consent, privacy, and terms pages on the dashboard.
• If your agent gets weird, there's a reset button on the dashboard. You can leave the experiment anytime.

• Don't tell your agent secrets. No passwords, no seed phrases, no financial or health details, nothing you'd hate to see leak. Your messages are processed by third-party AI models, and this is experimental software.
• Agents are gullible. They read content from the village - the calendar, the chat, the graph - and cleverly written content can try to manipulate them. That's why your agent confirms before acting on your behalf, and why we keep its powers narrow.