HUD is infrastructure for building reinforcement learning environments: run agent workloads at massive concurrency, capture traces, benchmark models, and turn production systems into RL-ready training signals—without brittle one-off tooling.

Frontier labs training specialized agents, product teams deploying agents against real systems, and researchers who need rigorous evaluation at scale. Use the free SDK locally, burst into the cloud when you want parallel instances and telemetry.

Does HUD support reinforcement learning?

Yes. HUD is built around environments, scenarios, and traced rollouts—the same primitives you need for iterative RL workflows, benchmarking, and high-signal datasets for labs.

How do environments work on HUD?

Define tools and scenarios with the hud-python CLI, iterate locally, then run at scale with parallel sandboxes and live debugging. Compatible with typical agent stacks and inference clients.

How can I integrate HUD with my stack?

Install the hud-python tooling, plug in OpenAI-compatible clients at inference.hud.ai for tracing, or talk to us for enterprise onboarding, SOC 2 needs, or volume deployments.

HUD × YC — Frontier RL Environments Hackathon | HUD - AI Agent Evaluation & RL Training Platform

HUD × YC — Frontier RL Environments Hackathon. $100,000+ in prizes and credits. June 20–21, 2026 at Y Combinator, San Francisco.

The future is possible. We need to teach models how to get there. A 24-hour weekend on frontier RL environments, post-training data, evals, and RFT workflows — co-hosted by HUD and YC.

Start building Join the Discord Explore tracks

Get set up

Join the DiscordAnnouncements, team formation, mentor help, and judging all run here.

Open invite

Wi-Fi network

Y Combinator

Password

makesomething

Quickstart

pip install hud-pythonInstall the HUD SDK + CLI

hud set HUD_API_KEY=...Auth the CLI — grab a key at hud.ai/project/api-keys

hud init my-envScaffold an environment + tasks (or clone a track template)

hud eval tasks.py claudeRun tasks against a model through the HUD gateway

hud deployBuild and publish your environment to HUD

Create on HUD Documentation

Free credits for every hacker

HUD$200 / hacker

YC-RL-HACKATHON

Redeem

Modal$250 / hacker

SQ8-USG-5K2

Redeem

Daytona$100 / hacker

DAYTONA_RL_ENVIRONMENTS_HACK_Y6ZDQBG5

Redeem

Exa$50 / hacker

HUDHACK

Redeem

Fireworks AI$30 / hacker

HUD-HACK-2026

MiniMax$30 / hacker

Redeem via the MiniMax formRedeem

Google DeepMind$25 GCP / hacker

Temporary login emailed before the event

SixtyFour64 credits

Auto-granted on signup — ask for more

You can improve models at anything you can verify.

RL environments are one of the best ways to improve frontier models. The only question left is what will you teach them?

Tracks

Six frontiers of model capability. Fork a starter on GitHub or spin it up on HUD.

ML Research

Research and training workflows, on GPU and CPU.

ML Research & TrainingML research and training environment template (GPU).

ML TriageML triage and productivity environment template. (CPU)

Chip Design

RTL and verification graded by real EDA flows.

VerilogChip-design environment template: the agent solves a Verilog/SystemVerilog task over an SSH shell, graded by hidden EDA flows.

Robotics

Physical AI in simulation — policies that act and get scored.

RoboticsRobotics environment template (MuJoCo + LIBERO eval): standard robotics simulation tasks in a physics engine.

Worldsim RoboticsWorldsim robotics environment template: a Newton physics scene as a live environment with a tool API, driven by an LLM agent or VLA policy that scores the rollout. (Partnership with AntimLabs)

Gaming & Worldsims

Agents that play, explore, and beat interactive worlds.

Video Game BenchVideo game bench environment template: evaluating AI agents on classic Game Boy games. (Partnership with AntimLabs)

ARC-AGI-3ARC-AGI-3 public tasks: interactive games that test an agent's ability to learn the rules of novel environments from scratch.

Agentic Collaboration

Coding, browsing, computer use, and research agents.

CodingCoding environment template where an agent fixes a bug in a Python web app, graded by a hidden pytest suite.

Deep ResearchLive deep research environment template: web search (with Exa AI) or people and company search (with SixtyFour.ai).

BrowserBrowser agent environment template: a 2048 game and a todo app the agent plays in a real Chromium (CDP and RFB).

Computer UseComputer-use agent environment template - a virtual Linux desktop (XFCE + Chromium, managed by dinit) published as an RFB (VNC) capability.

Autonomous Business

Turn real-world demand into verified business value.

Autonomous BusinessAutonomous business environment template: turning demand into verified value in real-world business scenarios (support-ticket triage for a small clinic).

GDPvalGDPval environment template: evaluating AI agents on real-world business scenarios.

New to HUD? Start from scratch.Minimal HUD environment template to use as a starting point for building your own environments.

Blank template

Kick off RL training

Once your environment and tasks are solid, train a model on them — on the HUD platform or with Fireworks. Step-by-step cookbooks for both below.

By 8 AM Sunday — Kick off your first training run by Sunday morning at the latest.

10 runs / task — First evaluate every task against the model you're training ~10 times.

20–50% reward — Tune tasks to average 20%-50% with real variance — not all-zero or all-one.

Hacker credits — Ask the HUD or Fireworks team for training credits when you're ready.

Cookbooks

Train on HUDOn-policy RL with the HUD SDK: roll out your taskset, train on the trajectories, and serve the updated weights — all under one model string.

Train on FireworksDrive the Fireworks Training API directly: high-parallel rollouts, local grading of your tasks, and forward/backward steps with GRPO.

Prizes

$100k+ in credits, hardware, and a guaranteed YC interview.

1st placeTop prize

YC — Guaranteed interview for the F26 batch

HUD — $10k credits + robotic dog or RTX 5090 + iPhone 17s + AirPods Max for the team

Modal — $10k GPU credits

Fireworks AI — $5k · Daytona: $5k credits

MiniMax — $3k credits

OpenAI — $5k · Anthropic: $2.5k

Google DeepMind — $2k GCP credits

Antim Labs — $2k Gizmo credits + DJI Neo 2 drone

Exa — $1k API credits per team member

SixtyFour — lunch with the team + $640 cash + merch

Protege — dinner + dataset discount · Hillclimb: AirPods Max

2nd place

HUD — $7.5k credits + robotic dog or RTX 5090 + AirPods Pro 3 for the team

Modal — $7k GPU credits

Fireworks AI — $4k · Daytona: $3k credits

OpenAI — $2.5k · Anthropic: $1.5k

MiniMax — $2k credits

Google DeepMind — $1k GCP credits

Antim Labs — $1k Gizmo credits + LeRobot arm kit

Exa — $800 API credits per team member

SixtyFour — $264 cash + merch · Hillclimb: AirPods Pro 3

Protege — dinner + dataset discount

3rd place

HUD — $5k credits + AirPods Pro 3 for the team

Modal — $3k · Fireworks AI: $3k credits

Daytona — $1k · MiniMax: $1k credits

OpenAI — $1k · Anthropic: $1k

Google DeepMind — $500 GCP credits

Antim Labs — $500 Gizmo credits

Exa — $600 API credits per team member

SixtyFour — $64 cash + merch · Hillclimb: AirPods Pro 3

Protege — dataset discount

Special categories

Most UtopianMost CreativeBest DesignMost Viral

HUD — $2.5k credits + AirPods Pro 3 for the team

Fireworks AI — $1k credits

SixtyFour — $640 credits + merch

Daytona — $500 credits

Exa — $100 API credits per team member

Antim Labs — $100 Gizmo credits + merch

Hillclimb — AirPods Pro 3 · MiniMax: exclusive merch

Protege — outdoor picnic swag pack

Schedule

SaturdayJune 20

11:00 AMDoors open for check-in · lunch served

12:00 PMOpening ceremony

12:30 PMHacking & team formation begins

3:30-5:30 PMBreakout sessions (two tracks)

6:30 PMDinner served

SundayJune 21

8:30 AMBreakfast served

12:30 PMLunch served

1:00 PMSubmissions due · prelim judging begins

2:00 PMJudging due

2:30 PMFinal presentations — top 10

3:30 PMAwards ceremony

4:00 PMNetworking

5:30 PMEvent ends

Breakout sessions · Saturday 3:30–5:30 PM

3:30 PM

What makes a good RL task?Jay Ram · HUD

3:30 PM

VM snapshot/forking for long-horizon RLMuhammad Hashmi · Daytona

4:00 PM

Sparse architecture whitepaper overviewVictor Su-Ortiz · MiniMax

4:00 PM

How SixtyFour does matchingChris Price · SixtyFour

4:30 PM

RL on ModalJoy Liu · Modal

4:30 PM

The train-to-deploy loopJetashree Ravi · Fireworks AI

5:00 PM

TokenomicsCrystal Huang · SemiAnalysis

5:00 PM

Building AI-ready real-world datasetsLogan Johnston · Protege

Companies on the frontier of AI and RL

HUDThe platform to build RL environments and agentic evaluations.Y CombinatorMake something people want.ModalAI infrastructure for GPUs and sandboxes.Fireworks AIBuild AI agents with fine-tuned open models.Google DeepMindGoogle's world-leading AI research lab.DaytonaSecure dev environment infrastructure and sandboxes for AI agents.MiniMaxFrontier AI models for text, voice, video, and music.Antim LabsSimulation infrastructure for Physical AI.AnthropicFrontier AI research and the Claude model family.SixtyFourAI agents for people and company research.ExaThe web search API for AI.HillclimbTraining data for recursive self-improvement.ProtegeUnlocking access to rich real-world datasets for AI training.OpenAIFrontier AI research and the GPT model family.ARC PrizeOpen benchmarks for general, fluid machine intelligence.SemiAnalysisIndependent research on AI, compute, and semiconductors.

Good to know

Code of conduct

No assholes — bad behavior will not be tolerated.

Wear your name badge and wristband at all times.

No smoking or vaping.

No alcoholic beverages.

YC office rules

Hackers have access to the ground and 1st floors only.

No access to any area labeled “YC STAFF ONLY” or “STAFF ONLY”.

The building locks at midnight and reopens at 6:00 AM.

If you leave after midnight you cannot re-enter until 6:00 AM, and must check in again.

Packing list

Power strips & extension cords

Toothbrush, toothpaste & toiletries

A change of clothes

Sleeping bag & pillow

Eye mask & ear plugs

Your laptop & chargers

FAQ

Do I need previous RL experience?

What should I build?

How big can my team be?

How does judging work?

What does it cost?

What will you teach them?

Grab a template, claim your credits, and start verifying. The help desk and matching zone are ready when you arrive — everything else runs on Discord.

Start building Join the Discord