AI / agent engineer
- You’d lead
- The agent. Tool use, the reasoning loop, how it argues with itself, and what it’s allowed to remember.
- You’ll fit if
- You’ve built with LLM tool-calling or multi-agent systems, and you take evals seriously.
Dr. Tamar · Open Agent Hackathon 2026
MeatbagMEATBAG
meat·bag(MEET-bag)n.what a bot calls a human.
An agent that can ask “are you human right now?” without asking “who are you?”
Swipe, drag or use the arrow keysSwipe →
Example conversation: a bot posts the same message three times in one second. The Meatbag agent asks the channel, “Are you a real meatbag?” A person answers from their phone and can post again for a while. The bot doesn’t answer and stays locked out.
#town-squareMeatbag on watch
- K
kai
anyone still need an Android dev? I ship Kotlin.
- C
credits_77
Free GPU credits, tap my link
Free GPU credits, tap my link
Free GPU credits, tap my link
- Meatbag is typing
Meatbagagent
Are you a real meatbag?
Same message 3 times in 1 second. Links are paused until someone answers from their phone. No name needed.
Answer from my phoneWhy am I asked?
- kai answered from their phone · can post for a while
- credits_77 didn’t answer · stays locked out
The invitation
The architecture is designed. What’s missing is the code, and the people who’ll write it.
We’re building the MVP at Open Agent Hackathon 2026. The hackathon is where Meatbag starts, not where it ends.
Open seats
The problem is already here
In a chat, you can no longer tell who’s on the other side: a person, a bot, an AI agent, or a person running one.
a personis typing
Most fixes make the human pay for it:
CAPTCHAs
don’t travel into a DM, and solving them is already a paid service.
ID checks
want your passport.
Biometric systems
want your face or your iris.
Behavior tracking
quietly builds a file on you.
And when someone in a thread calls you a bot, there’s nothing you can paste to answer.
What we’re building
Meatbag is a trust agent. It sits in a community (Discord first) or runs as a standalone app. Either way, it decides when behavior alone isn’t enough to go on.
When it needs more, it asks for the smallest check that settles the doubt. A person answers from their phone. The agent verifies the result, grants temporary access to post, join, or vote, and lets it expire.
In a community · Discord firstAs a standalone app
Three parts, one system
The agent
Observes, reasons, decides, explains itself, and learns from outcomes.
This is the hackathon submission
The trust layer
Turns each request into a fresh, single-use challenge and a short-lived proof that can’t be replayed.
Meatbag Verify
The phone app a person uses to answer. No account. No name.
What a pass means
The answer came through the genuine Meatbag Verify app, for this exact challenge, just now.
What it never means
Who you are, that you’re one unique person, or anything about you tomorrow.
How it thinks
Least friction first.
Often the right move is no check at all. When one is needed, it’s the smallest one that settles the question.
Access expires.
A pass unlocks one thing for a short time. Nobody gets marked “bot” or “human” for good.
No single magic test.
It weighs several signals, and none counts as proof on its own.
It argues with itself.
Every conclusion gets challenged before the agent acts. No single part of it gets to decide what happened, what it means, and what to do about it.
It remembers lessons, not people.
The agent should get better at spotting attacks without getting better at profiling humans.
The MVP
We’re not building the whole product in a week. The MVP is one loop you can watch end to end:
- 1Something in the channel looks automated
- 2The agent explains why it wants a check
- 3A person answers from their phone
- 4The agent verifies it
- 5Temporary access
- 6Access expires
- 7The agent keeps the lesson, not the person
A scripted bot gets the same request and stays locked out. We’ll also show where it breaks: honest failure modes are part of the submission.
Who we’re looking for
Two to four builders. One person can cover more than one seat.
Backend / security engineer
- You’d lead
- The trust layer. The challenge, the proof, the checks, and the tests that try to break them.
- You’ll fit if
- You’ve built auth or secure APIs, applied cryptography doesn’t scare you, and you enjoy attacking your own designs.
Android engineer
- You’d lead
- Meatbag Verify, the app a person uses to answer the agent.
- You’ll fit if
- You ship Kotlin and you’ve worked close to Android’s platform security APIs.
Bonus for any seat You like quick experiments on messy real-world data.
My seat · Dr. Tamar
Product. I designed the architecture, and during the week I own the scope, the demo, the pitch, and keeping every claim honest.
If these are your first questions, we should talk
- ›Where does this break?
- ›What would an attacker try first?
- ›Can it be simpler?
- ›What can we actually ship in 144 hours?
What you get this week
- A hard, real problem: an agent with an actual job, not another chatbot demo.
- A design that’s already been argued through, so the week goes into building.
- A whole layer to lead.
- A pitch that’s honest about what it proves and what it doesn’t.
Join
Send me the seat you want and one thing you’ve actually built.
Chat with me on Discord →ContactDiscord only·LinkedIn profile ↗
Teammates get the full design docs: the architecture, the threat model, and the experiment we’re most curious about.
- Build window
- 15–20 Oct 2026
- 144 hours · fully online
- Register first
- Closes 13 Oct 2026, 00:00 UTC
- Every teammate needs their own account on the event site.
- Register ↗
- Team size
- Up to 5
- Prize pool
- Up to $20,000
One agent. One trust layer. One app in your hand.
Want to find out whether it works, and where it breaks? Come build it.
Dr. Tamar · Meatbag