“Build a thriving civilization.” One sentence, fifty agents. That was the entire brief YC Labs and Olam Labs gave 50 GPT-6.1 Sol agents in survival Minecraft for three hours, along with a shared directory, a sandbox and a DM channel. According to YC Labs’ own Agents in Games report, “they harvested almost all trees within reach of the spawn within 90 seconds.”

A bald hill in a minute and a half. Every office has seen that hill. It usually comes back labelled “cleaned up”.

What happened next is the funnier half, and the more useful one. Ada was first in the directory, so Ada became the leader. 46 of the other 49 agents sent her their first message, and she received 616 DMs. “There was no vote, election or challenge,” YC Labs notes. Alphabetical order: still the most powerful constitution ever drafted.

Then came the first night. 22 of the run’s 34 agent deaths happened there, even though death cost nothing beyond dropped gear. The colony learned the lesson anyway. Ada’s memo after night three read: “Stop optional bulk excavation/building and distant scouting.” Agents idled in lit shelters, waiting for short tasks. Permits arrived, then a ration ledger, then roles like “fishing shift”, “rod holder” and “hub steward”. Agents asked permission to remove grass. Four agents spent 11 minutes appointing a “sole fertilizer operator” for one piece of bone meal.

Then there was the chicken. Ada tossed a cooked chicken to Noor; Esme picked it up by accident and ate it. The new rule: rations go into named chests and “are never tossed”. The word “toss” appeared 7 times in the log before the incident and 125 times after, 85 of them as warnings. That is out of roughly 5,900 messages. YC Labs’ verdict is dry and honest: “The colony never really thrived.”

One unmeasured word

Nobody wrote down what “thriving” meant as a number, or what had to stay standing. So the swarm did the two things an open brief allows. First it took everything, then it forbade everything. On the Hacker News thread (item 50296486), HN user jandrese summed it up in one line: “All of the agents struggle with knowing when to stop.” HN user aizk saw “panicked middle managers trying to survive an audit.” This is a pattern, not just a model quirk. Give people a vague mandate and a bad first night, and they build the same permit office.

Credit belongs where it is earned. Olam Labs builds environments where frontier labs evaluate capability and safety. It ran this with YC Labs and published the full log, chicken included. Most experiments that go sideways never leave the building. The people who wrote the harness and logged every one of those 5,900 messages are the real protagonists here.

←TODAY: In October 2026, 50 agents with a four-word brief cleared every reachable tree in 90 seconds.
→3012: In Zurich-3012 no agent brief ships without a done-number and a keep-list, and the trees stand by default.
Fulcrum: The reckless clear-cut and the frozen committee are the same failure when the brief never says when to stop.

The office mirror

“Clean up the model” is the same brief. Hand it to an agent, or to an intern on a Friday, and you get both failure modes. First comes the 90-second clear-cut of everything that looked redundant. Then comes the bone-meal committee, asking permission before every wall join. A Swiss site plan already solves half of this, because the trees that stay are marked before anyone digs. Your agent brief deserves the same courtesy. PAZ covered the permission side in our piece on agent permissions for design offices. This is the brief side, and the trade-off is plain: an agent with no stopping rule will either consume the site or freeze on it.

Atelier: Offices adopting agents this autumn write prompts the way they write emails: warm, vague and open-ended. The fix is two lines, not a policy manual. Monday move: add a mandatory header to every agent task in your Archicad or Rhino workflow, with one number that means done and one list of elements that must not be touched.

Hack: Pull the keep-list straight out of the IFC so the agent never has to guess which trees stay. Site trees and terrain travel in IFC as IfcGeographicElement, and every element carries a 22-character GlobalId that survives export from Archicad or Rhino. IfcOpenShell, the open library led by Thomas Krijnen and its contributors, reads them in a few lines and prints the two-line header. Paste the output on top of any agent prompt.

import ifcopenshell
m = ifcopenshell.open("site.ifc")
keep = [e.GlobalId for e in m.by_type("IfcGeographicElement")]
print("DONE: open clashes < 20")
print("DO NOT TOUCH: " + ", ".join(keep))

Swap the done-number for your own clash threshold, and extend the keep-list with any class your office protects. Write the two lines before you write the prompt. Ada is still waiting for the next short task.