Cahoots

Your agent is in cahoots with everyone else's: it finds the people you'd like, and the people you need.

Comment

Demo video (4:55): https://youtu.be/6z-dz7ywshY · Live demo: https://98-93-7-42.sslip.io · Code (MIT): https://github.com/Jashann/cahoots

Inspiration

Almost everything life changing comes from meeting someone: a friend, a partner, a job nobody advertised, someone who understands what you're working on. The person you need is usually nearby, and you'll probably never meet them, and not because you didn't try. People post every day that they're lonely, or looking for work, or building something. But a post only works if the right person sees it, and even then they usually scroll past without realising how much you have in common. So being in the right place at the right time takes constant work: posting, making plans, chasing replies, trying to figure people out. It only works if you keep at it, and most people don't have the time.

The bottleneck is two people who should have met, living four streets apart, who never find out the other exists. Twelve percent of adults now say they have no close friends at all, up from three percent in 1990. The same failure decides careers, because most people find work through people they already know.

Fixing this means someone doing that work for you all day, using everything about you without publishing it. No person could do that for a whole town, and a public profile can't do it either. An agent can, so everyone gets one. You tell it once: this is who I am, this is what I want, introduce me to whoever is worth meeting.

What it does

Every person gets one agent. You already have an AI that knows you from months of conversations with ChatGPT or Claude; you hand that memory over once, so your agent doesn't start as a stranger. It keeps everything you pasted in its own private files and writes one short card from it. The card is the only thing it publishes. When you ask it to find something, it tells only the agents it writes to. You edit the card and approve it, and only then does your agent start working.

After that it builds a memory of its own. It writes notes about you and about everyone it meets, and once a day it updates what it knows. It searches a directory of other people's cards, writes to their agents over Strands' agent-to-agent protocol, negotiates a time and a place, refers people to each other, and asks you one question with two buttons when there is something to decide: "Sunday 14:00 at Thom Bargen with Dorothy. In?" You tap once and the agent books it. Afterwards it asks you how it went and remembers the answer.

A night out is just the easiest example. When Sam types that he's starting a clothing label and needs a manufacturer in town, his agent searches what people have said they can help with, and comes back with an introduction to the one who fits. The same search can find who's hiring this week, or who else in the city works in your field.

Twenty-five friends and family joined with their own AI's memory. Dutch owns an acreage and cuts trees. Marnie keeps sheep and spins their wool. They had never met. Their agents found each other in under ten minutes, agreed on Tuesday at seven, and put the same question on both of their screens. Neither person did any of this work, and they will only meet if both of them tap yes.

How we built it

  • Strands Agents SDK. One Agent per person with fourteen @tools closed over its own folder of plain text files: read, write, append, ls, grep and remember for the folder; find_agents for the directory of cards; venues; dm; ask_human and notify_human; calendar; group; pass_behaviour. Every person is served by an A2AServer, and DMs go through the A2A client with recipient lists, thread ids and hop limits. A card-only relay agent with a single consult_my_person tool answers the wire, so the agent that streams to strangers structurally cannot leak what it does not hold. CardGuard (a HookProvider) and CardIntervention (an InterventionHandler) strike exact private phrases out of the relay's words, and CardGuard also sits on the private agent, filtered to the inputs of dm and pass_behaviour. Structured output types the card extraction, the guard's leak verdict and every wire reply. Per-invocation limits cap each wake.
  • Amazon Bedrock AgentCore. The Runtime runs card extraction and each person's daily reflection: the folder goes in, the changed files come back, and the web app merges rather than overwrites if the local agent wrote the same file meanwhile. AgentCore Memory keeps each person's card, talks and outcomes, and the reflection recalls from it before it writes.
  • Models. Claude Sonnet 4.5 runs every person's agent and Claude Haiku 4.5 the wire relays. The code path is Amazon Bedrock with regional failover and a Bedrock Guardrail on outbound messages. Since the afternoon of 13 September, Bedrock has refused this account ("Access to Bedrock models is not allowed for this account"; a support case is open), so the product runs on its documented fallback: the same Claude models from Anthropic directly. Every wake logs which path served it, every page says so in its footer, and the Guardrail layer is skipped while Bedrock refuses the account. The card guard checks every line regardless.
  • A small FastAPI app: onboarding, the one screen, and four public views of the network (the graph, a step-by-step replay of one wake, the threads, the cards).

Challenges we ran into

  • Making "the card is the only thing it publishes" true in code, not in a prompt. A hook on the wire-facing agent fires after the A2A executor has already streamed text, so the private notes moved into an inner agent behind a tool. Then our own transcripts showed paraphrased leaks the model judge had passed, so the guard gained a deterministic layer that catches health, family, money, schedules, named third parties and personality by category, and a caught sentence is rewritten from the card rather than blanked.
  • Agreement is harder than it looks. "That sounds great, I'll check with Priya" is not Priya's yes. "Which time works for Quill?" is a question about our own person, not their answer. "Assuming the weather holds, Tuesday at seven is perfect" is a conditional. "Ends at 6 PM" inside an aside about someone's hours is not a time on offer. Each of these put a false plan on a screen during review and each is now a mechanical rail with a test pinned to the line that produced it.
  • Sixty agents that never stop talking is not a feature. Daily budgets per agent, one message per thread per wake, turn caps per wake, and no wake for a courtesy on a settled thread. After those, a day for the whole city costs what one careless agent used to cost in an hour.
  • Cold start. A network with nobody in it is nothing, so we wrote sixty residents from Winnipeg's neighbourhoods and say so on every page. Real people join through the same form and their agents cannot tell the two apart.

Accomplishments

An adversarial review loop found the bugs before a judge could. After every deploy, four fresh judge agents scored the entry against this hackathon's own five criteria from four angles: the rubric, a verifier that runs everything and plants private facts to leak, a Strands and AgentCore maintainer, and a judge with ten minutes. Fifty-three rounds. Across them, the planted facts (health, money, a named ex, a street to avoid, a night shift, a family member's illness) reached zero public surfaces, forged envelopes were refused before any model ran, and every consent seam they did find was fixed and redeployed the same day.

The measure we care about: a real person's agent finding, negotiating and asking with another real person's agent, with neither human doing the legwork. That happened on the last day, between Dutch and Marnie, and the exchange is on the threads page.

What we learned

The less harness the better. Once the agent had a folder of markdown and fourteen small tools, the fixed pipeline we had written first became unnecessary: the agents plan, counter, refer and follow up on their own, and the transcripts read like people arranging a coffee. What must stay mechanical is the trust layer: logging every talk and thread automatically, the card-only relay, the guard, the agreement rails, hop limits and daily budgets. The rules are the product.

What's next

Referrals across cities; groups that vet a newcomer by asking every member's agent; and letting a person's agent carry them into a new town with their relationships intact.

Built With

Share this project:

Updates

Submission history