M-AGENTS · June 7, 2026 · Fordham Lincoln Center · NYC Tech Week · a16z

One problem.
Four agents.
Your product.

You're given a real business crisis. Build a multi-agent system that addresses it — then ideate and design your own product that puts the solution in the hands of someone who needs it. The crisis is the prompt. The product is yours.

1 Day 4–6 Per Team 400+ Registered 2 Tracks $2,000 Cash
RSVP on Partiful ↗ Register on Devpost ↗
M² multi-agent hackathon brand mark

Built with

Trupeer Cognee LingCode PyMC Labs Geodo
The Challenge

Build agents that solve a real crisis. Build a product that solves it for real people.

M-AGENTS is a single-day build sprint. You receive a real business crisis — corrupted data or undetected fraud — and four hours to ship a working multi-agent pipeline. Then you turn that pipeline into a product your end user can actually operate. Judging happens live, science-fair style.

The Tracks

Pick a track, then ideate freely within it.

You define the product, the user, and the success condition in your Product Brief. The track is the problem space — not the spec.

Track 01 — Data Rescue
DATA
RESCUE

A manufacturer's data is corrupted four days before a regulatory audit — duplicates, unit conflicts, numbers that contradict each other. Build any product that helps an organization find, fix, and explain broken data.

End user archetype: the compliance officer who has never opened a database.

Track 02 — Fraud Watch
FRAUD
WATCH

A fraud ring stayed under every alert threshold — small transactions, circular flows, coordinated accounts. Build any product that helps a fraud team see what rules miss.

End user archetype: the analyst with three minutes per case.

Datasets hosted on Kaggle — optional. Use the benchmark and judges verify your findings against a hidden answer key for bonus credibility. Or bring your own data.
Track 01 dataset: View ↗  ·  Track 02 dataset: View ↗
Team Roles

Everyone has a job. Non-technical people aren't support.

Builder
BUILDER

Makes the agents run, wires Cognee, connects agent outputs to the product interface.

Designer
DESIGNER

Owns the Product Brief and everything the user sees. Step 0 and Step 5 are yours.

Domain Expert
DOMAIN EXPERT

Validates outputs, writes the Agent 4 narrative, and runs Geodo research on real-world entities.

Presenter
PRESENTER

Owns the demo script, the Trupeer recording, and the finalist stage presentation.

Every team needs at least one Builder and at least one other role. A team of one Builder and three sharp non-technical people can win this.
The Five Steps

A clear pipeline from brief to product.

00
Define
Write your Product Brief first.

One page before any code. Who's it for, what does it do (one sentence), what does success look like, what won't you build.

01
Find It
Agent 1 reads the data.

Reads the dataset. Finds what's wrong — duplicates, anomalies, suspicious patterns.

02
Rank It
Agent 2 prioritizes findings.

Sorts the findings. Worst first, with reasons for every ranking decision.

03
Act On It
Agent 3 fixes, flags, or escalates.

Takes action on each finding. Every action has a logged reason — "the model said so" fails judging.

04
Explain It
Agent 4 writes the summary.

Produces a narrative a human can read and sign. Downloadable from the product. Domain Expert owns this step.

05
Show It
Build the product. Demo as your user.

Build the product from Step 0. Demo it as your end user — not as an engineer explaining code.

Memory connects everything — each agent recalls the previous one's work through Cognee.

The Stack

Tools that power every team.

Mandatory tools are required for a valid submission. Recommended and optional tools are there when you need them.

LINGCODE.DEV
Recommended IDE

Design and customize your agents. Claude Code, Codex, and Gemini CLI built in. Web, CLI, or native Mac.

lingcode.dev →
COGNEE
Mandatory

Memory layer between all four agents. 14-day Cloud trial. Top prize: $200/mo credits. 2nd & 3rd: $35/mo credits.

cognee.ai →
Discord help room: Join Discord ↗
TRUPEER
Mandatory

Your 5-minute demo video. Required submission. 14-day trial included for all participants.

Start Trupeer trial ↗
GEODO
Mandatory

Domain Expert researches your product's real-world entities — customers, companies, market. Web platform, no code. 100 credits + 5-day Pro included.

geodo.ai →
Sign up at geodo.ai first, then: Request access ↗
KAGGLE
Data

Both track datasets. Optional — but using the benchmark earns bonus credibility: judges verify your findings against a hidden answer key.

Track 01: View ↗
Track 02: View ↗
PYMC OPEN-SOURCE STACK
Optional · Special Prize
PyMC

Probabilistic reasoning + Bayesian modeling.

github.com/pymc-devs →
PyMC-Marketing

Media mix modeling + budget optimization.

github.com/pymc-labs/pymc-marketing →
Decision Hub

15,420+ validated agent skills. Chat with it to discover what's possible before writing your brief.

hub.decision.ai →
Daimon

Data-scientist agent in the PyMC Labs Discord.

Join Discord ↗
Decision Lab

Harness for agentic data science.

github.com/pymc-labs/decision-lab →
Before writing your Step 0 brief, ask Decision Hub what agent skills already exist — it may save you hours.
Register

Three steps to lock your spot.

01
RSVP

Reserve your seat at the event on Partiful.

RSVP on Partiful ↗
02
REGISTER YOUR TEAM

Two registrations — Devpost for your submission, the form so prizes reach the right team.

Devpost ↗
Team verification form: Verify team ↗
03
REPORT ISSUES DAY-OF

Something's not working on the day? Use the issue form.

Report an issue ↗
Schedule

June 7, 2026 — Fordham Lincoln Center

9:30 AM
Doors open · Breakfast · Networking
10:05 AM
Lightning Talk — Christian Luhmann, COO, PyMC Labs (virtual)
10:17 AM
Lightning Talk — Dave Nielsen, Head of DevRel, Cognee
10:29 AM
Lightning Talk — Atin Woodard, Founder, Stage 11 Agentics
10:45 AM
Team formation · Role cards · Challenge reveal
11:00 AM
Build sprint begins — Step 0 Product Brief first
1:00 PM
Lunch
4:00 PM
Science fair — judges walk the room
5:00 PM
Submissions close on Devpost — no extensions
5:00–5:20 PM
Judges deliberate · Three finalists selected
5:20 PM
Finalists announced
5:25–6:00 PM
Finalist demos · Special prize demos
6:00 PM
Awards · Job opportunities · Close
Rules

Ten rules. No exceptions.

R01
Minimum 4 agents with real handoffs — not one LLM call in a loop.
R02
Cognee is the memory layer — every agent reads from and writes to it.
R03
Trupeer demo video is a mandatory submission. No video = disqualified.
R04
Geodo research is mandatory, done by the Domain Expert, web platform only.
R05
Product Brief (Step 0) is required, submitted, and judged against your own product.
R06
Data: bring your own or use the Kaggle benchmark — benchmark earns bonus verification.
R07
Every agent decision must have a visible reason — "the model said so" fails judging.
R08
Agent 4's summary must be downloadable from the product.
R09
Teams of 4–6. At least one Builder and one other role. No solo submissions.
R10
Submit on Devpost by 5:00 PM — no extensions.
Judging

Five criteria. 25 points maximum.

1
Agents that work

Ran on real data. Outputs aren't hardcoded.

5 PTS
2
Real collaboration

Agent N+1 demonstrably used what Agent N found, via Cognee.

5 PTS
3
Matches your brief

Judged against YOUR OWN Step 0 success condition.

5 PTS
4
End user can use it

A judge operates the product cold during the science fair.

5 PTS
5
Explainable

Every decision has a visible reason a human can follow.

5 PTS

Minimum 15 points to qualify for finalist consideration. In-person judges walk the science fair 4–5 PM. Virtual judges review Devpost from 5 PM. Three finalists demo live.

Prizes

$2,000 cash + tools worth more.

1st Place
$1,500
Cash prize
2nd Place
$200
Cash prize
3rd Place
$100
Cash prize
Best Use of Trupeer
$200 cash prize
Geodo Top Team
3 Pro accounts (~$3,000 value)
Cognee
$200/mo cloud credits (1st) · $35/mo credits (2nd & 3rd)
PyMC Special Prize
Course seat (~$2,000 value)
Everyone gets: 14-day Trupeer trial · 14-day Cognee Cloud trial · 100 Geodo credits + 5-day Pro access.
Job opportunities announced in the room — SWE + GTM roles from sponsor companies.
How-To

Get up to speed before the day.

Recording your demo with Trupeer
Recording your demo with Trupeer

Record, narrate, and export your 5-minute required demo video.

Building agents in LingCode.dev
Building agents in LingCode.dev

Use Claude Code, Codex, or Gemini CLI to scaffold and wire your agent pipeline.

FAQ

Common questions.

Yes. Designer, Domain Expert, and Presenter are core roles — not support. Steps 0, 4, and 5 and all Geodo research belong to non-technical teammates. A team with one Builder and three sharp non-technical people can win.

No — both live on Kaggle and they're optional. Bring your own data if you prefer. Using the benchmark earns bonus verification: judges check your findings against a hidden answer key for extra credibility points.

Yes — bring your own key for OpenAI, Anthropic, or Groq. Groq has a free tier with no card required. No API keys are provided at the event.

Your Devpost entry must include: Product Brief PDF, GitHub repo link, Trupeer video URL, track selection, and a written description of your product. All items required — partial submissions are not accepted.

Science-fair scores (4–5 PM, in-person judges) plus submission review on Devpost (virtual judges, from 5 PM) → the top 3 teams demo live at 5:25 PM. Minimum 15 of 25 points to qualify.