Warehouse Challenge · 2026-06-27 · winners pov

The REFUGIO hackathon challenge!

On the last Saturday of June 2026, 48 people were shut in a room in Madrid for four hours. The organizer, a Twitter nerd with a Milhouse pfp (@XMihura), had picked them by hand out of more than 330 applications and handed all sixteen teams the same problem: run an Avazon warehouse. Beff Jezos was in dire need of our help, and needed to optimize his warehouse in only 4 hours. Milhouse wrote a paper about it afterwards, but I'm here to tell the story from the other side: what the competition looked like from the POV of the team that won.

Our team was TEAM 03. We took the lead at 20 minutes in, lost it only a few minutes later, and spent the rest of the challenge in second and third places, fighting to win it back at 12:47. We managed to claim our 1st spot back by beating the record by a single delivery, three times over. Before I tell you our story, let me run you through the challenge in more detail first.

The scoring system

We have to work with a 52×52 grid, 96 robots docked around the edge, and 960 shelves in the middle, with a fixed amount of time to deliver as many packages as possible. Points only go to the first team to push the all-time record higher, with each newly claimed delivery count above 100.

Breaking the record is the only thing that lets you collect the bounties on every delivery count above the old record, up to and including yours. If you only match the existing record you get nothing at all. So a big jump is worth a fortune: taking the record from 398 to 610 collects 212 bounties in one go, worth 85,754 points. And being fast is promoted; sitting on a good solution is risky, because once another team claims a delivery count, nobody else can claim it ever again. Under these rules, eight teams finished the day on 895 deliveries or more with zero points.

My mental map

Once the scoring was disclosed, I quickly understood the problem as three optimization vectors. First, the robot policy — what each of the 96 robots decides to do next: the algorithm, the routing, the collisions. Second, the shelf organization: where each of the 960 shelves actually sit. Third, and the least glamorous, how fast you can iterate: we were given out the verification code that runs a candidate and hands back a number, but no one said it's fully optimized, and faster iterations means more iterations, which means better solutions, faster.

Before the competition started, Mihura added a fourth vector to optimize: exploits were fair game. If the simulator let you do it, it counted. There were some safety guardrails, like an LLM that scanned for malicious code, but anything that passed that was ok. This was a big optimization vector, but a hard one to get with LLMs, as they are not happy to investigate and share where the setup leaks, even if it would be fair game.

In a traditional hackathon, with only 4 hours on the clock, you would really have to prioritize which optimization vector is the most advantageous at each point in time. You can't clone yourself and explore all 4 at the same time. With coding agents, you very well can, and that's exactly what we did.

Phase I · 10:06 – 10:31 · the brain

Just 30 minutes in, the day was over for most teams

Almost every point worth winning was gone inside 25 minutes. The third solution was ours; it basically won us the whole thing

Phase II · 11:00 – 13:58 · the floor

As the robot policy saturated, the fight moved to the floor

With the brain commoditized and copied, teams started experimenting with the floor. The initial shelf layout branched into 20 different ones

1TEAM 0288,683

2TEAM 0887,165

3TEAM 03 · us85,754

5TEAM 107,938

Those were the three teams that would fight for the prize from here, with one exception: TEAM 10, fifth on 7,938, would finish second

Rewriting a routing policy is a coding problem, and thus the coding agents were amazing at it. However, when that optimization vector stopped giving results, redrawing a warehouse was the next one. It is, however a visual optimization problem: where should the aisles run, where do robots pile up, what does this floor look like when it jams. The SOTA models at the time were way worse at that than coding, which is the same weakness ARC-AGI-3 surfaced. Newer models like Astra seem to have unlocked it by representing the visual as something closer to a set of equations than a picture

Phase III · 13:27 – 13:58 · the finale

In the last half hour, TEAM 10 found another 78 deliveries

They ended the day with the best warehouse in the room by a distance. It was worth second place.

We didn't get the best solution… but we got the victory!

All the teams made good delivery numbers, but only half of them were fast enough to get any points at all.

1,008

TEAM 10 demolished the existing record last minute by overfitting the shelf organization to the 3 evaluation seeds.

record-breaks by us, more than any other team. TEAM 10 managed 4, nobody else more than 2.

teams finished on 895 deliveries or more and scored nothing.

Submission log

Our 610 at 10:22 took the record from 398 in one jump, and gave us the biggest leverage: 85,754 points. However, every single time we beat the record was decisive. In the end, a single additional delivery by TEAM 10 could've crushed our dreams an taken the victory from us. Every tiny, unglamorous record-break in the afternoon was ultimately essential to our victory. And I think that's beautiful. Thanks for getting this far!