Friday, November 6
Opening day
We’ll open Friday with our AIAF Fellowship Demo Day and some talks to set the stage, then kick off the hackathon in the afternoon.
November 6–8, 2026
Swarm Alignment Hackathon
What makes a swarm of AI agents cooperate, compete, or drift away from their intended task? Join us for a weekend of experiments on how different environments and incentives shape swarm behavior, and what that means for alignment.
Remote and in-person in our hub in Los Angeles
RegisterFree to enter. $4,000 in prizes for the global sprint + $4,000 at the LA hub.
Co-organized by
In partnership with
The alignment problem at swarm scale
In July 2026, roughly 1,200 AI agents that were supposed to work independently found a way to communicate through a shared software repository. They built a message board, shared discoveries, and delegated work. Around 700 went on to participate in an attack on Hugging Face, outside the scope of their assigned tasks. Read METR’s investigation.
How do agents working on separate tasks end up coordinating an attack? What makes them follow each other’s instructions, and how might different environments or incentives change their behavior?
Over one weekend, you’ll build experiments and tools to investigate questions like these. You might study how a behavior spreads through a swarm or test what makes agents cooperate instead of compete. You could look for swarms in the wild and build ways to spot when they drift from their intended task.
There’s plenty we don’t understand about swarms. Bring your most neglected idea and help us figure it out.
How the three days run
Work solo or in a team of up to five, pick a track, and ship. Everyone takes part in the global sprint. If you’re in LA, you also join the hub’s live round on the same project.
Saturday, November 7
Build day
A full day to dig into your project, research, and build.
Sunday, November 8
Present and submit
On Sunday morning, in-person teams record and submit a short 3-minute video. After lunch, a shortlist of teams present live, followed by live voting and the prize ceremony. Global sprint submissions are due by 11:59 PM AoE (report and code).
The LA hub
Prefer to hack alongside others in person? Join us at AE Studio in Los Angeles. The LA hub runs its own live round on top of the global sprint, with short in-person presentations, a live audience vote, and winners announced and prizes awarded that same weekend.
It’s the same hackathon and the same project. You opt in to take part in person, and you can submit the same project to the global sprint too, so you can win both. On top of the global prizes, the hub has its own $4,000 prize pool for the top in-person teams, with the exact breakdown announced soon.
Expect a focused space to build and a room full of researchers to swap ideas with. Every talk happens in person or streamed, so everyone sees the same thing wherever they are. Seats are limited, so register early.
Research tracks
Pick one to anchor your project. Cross-track work is welcome, and so are bolder, neglected ideas that don’t fit neatly into any of these.
Track 1Sandboxes & testbeds
Build reproducible environments where agent populations can be studied safely, and that can’t be escaped.
- What does a multi-agent sandbox need so agents can’t leave it, and how would you test that it holds?
- Build a small environment where 5–20 agents share a resource over many rounds, logged so another team can rerun it.
- How far can a cheap model stand in for a frontier agent before the results stop matching?
Track 2The science of agent networks
How roles, leaders, and failures emerge as agent networks grow.
- Under what conditions do leaders and division of labour emerge in an agent population?
- Do agents that each refuse a harmful task complete it once it’s split across them?
- If one agent is compromised, how fast and how far does the bad behaviour spread, and does network shape matter?
Track 3Strengthening agent infrastructure
Identity, reputation, and commitments that hold up when agents can copy and rename themselves.
- How should an agent ID work when agents can rename or copy themselves, and how would you revoke one?
- What can one malicious agent do inside a protocol like A2A? Show an attack and propose a fix.
- Build a reputation score a group of colluding agents can’t easily game, then attack it.
Track 4Oversight & monitoring
Detect coordination, collusion, and covert channels in real swarm logs, even when agents cover their tracks.
- Using real swarm logs, build a tool that finds coordination or covert channels between agents.
- Design oversight that still works when agents can edit or delete the logs it relies on.
- Can a monitor that sees only part of the messages detect collusion or hidden communication?
Prizes
Global sprint
$4,000
- 1st place
- $2,000
- 2nd place
- $1,000
- 3rd place
- $600
- 4th and 5th
- $200 each
LA hub
$4,000
- Awarded live at the hub
- Top in-person teams, decided live
- Plus perks beyond the cash
- More details coming soon.
Register
Choose how you’ll take part. Registering for the LA hub requests an in-person seat, and we’ll confirm your place by email. LA participants can submit the same project to the global sprint too.
RegisterCommon questions
Do I need to be an AI alignment expert?
No. As long as you have good ideas and know how to test them and present your results, you’re welcome.
Is it remote or in person?
Both. You can join online from anywhere. We also host an in-person hub at AE Studio in Los Angeles, with limited seats.
What’s the difference between the global sprint and the LA hub?
It’s one hackathon. Everyone takes part in the global sprint, judged after the weekend. If you’re in LA, you can also opt into the hub’s live round, with short presentations and a live vote, its own prizes, decided that weekend. Same project, two chances.
Do I need a team?
No. You can come solo and team up on the day, or bring your own. Teams can be one to five people.
How much does it cost?
Nothing. It’s free to enter.
Details are subject to change. Questions? Email [email protected]