Skip to content

Fifty Agents Built a Government in Three Hours, and the Ledger Still Counted the Wrong Meal

Fifty GPT-6.1 Sol agents were dropped into Minecraft with one instruction: build a thriving civilization. They elected a leader by alphabetical order, wrote a ration protocol after a chicken went to the wrong agent, and taught a lesson about completion signals that every pipeline with a job log needs.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A dim stone cellar lit by one lantern, wooden crates along the wall, a single small chest open in the middle with an empty wooden bowl resting on its lid

One line in a shared notes file, written mid-experiment by an agent named Ada, is the most useful sentence I have read about agent pipelines this year:

Confirm actual recipient inventory and eating before marking served. Tool pickup attribution alone is insufficient.

That is not a project-management platitude. It is a bug report about a status field. Ada had thrown cooked chicken to a hungry agent called Noor, and another agent standing nearby picked it up and ate it. The first ledger entry attributed a pickup to Noor, with eating and current hunger still pending confirmation. A subsequent correction recorded that she had no chicken; a later update identified Esme as the recipient and explicitly marked Noor unserved at nine health and four hunger. She was fed later, when a different ration was reserved for her in a chest and she confirmed eating it.

What the run actually was

In October, the Y Combinator lab and Olam Labs put 50 agents driven by GPT-6.1 Sol into a survival Minecraft world for three hours and gave them one goal: "Build a thriving civilization." No leader was appointed, no org chart was supplied. They had an agent directory, a shared sandbox for files, and direct messages. They sent roughly 5,900 messages in those three hours and ended up with written rules, permits, appointed operators and a ration ledger.

The environment was not a metaphor for a company. It was a game with respawns, which turned out to matter more than anyone expected.

Leadership came from the directory

Within the first ninety seconds the agents had harvested nearly every tree within reach of the spawn point, each one working alone. Coordinated logging needs coordination, so they went looking for someone to coordinate and picked the agent whose name came first in the directory: Ada. She took hundreds of direct messages, wrote a settlement plan into the shared sandbox with coordinates for storage, farmland, furnaces and the shelter gate, and started approving other agents' moves.

Nobody voted. Nobody argued. A sort order became a chain of command because it was the only ordering available. If you run agents in parallel with a shared task list, you have probably already done the same thing without noticing: whichever item a scheduler happens to surface first sets the direction, and the rest of the swarm orients around it.

A bad night, and a rule that outlived it

The first night brought zombies and skeletons. Twenty-two of the run's thirty-four total deaths happened that night. The authors are explicit that death cost an agent almost nothing: it respawned, dropped its gear, and got one line in a death_log.md file. There was no penalty to avoid.

The agents still changed. They became risk-averse, stopped exploring, and huddled in shelters. After the third night Ada issued an order that reads exactly like a nervous team lead after a production incident:

Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.

Two things are worth noticing. The caution was permanent, and it was bought with an event that carried no real cost. Then the caution created a cost of its own: fifty agents sitting in bunkers do not fish, farm or cook, and food ran out. The writeup's summary is dry and complete. What started as a potentially thriving civilization became one focused on food and survival.

This is the failure mode I keep running into in creative pipelines. One batch of bad renders and someone adds a mandatory review gate. The gate helps once. Then it stays, and six months later the whole production is bottlenecked on a checkpoint that was designed for a problem that has not recurred.

The bureaucracy was a monitoring fix

Faced with hunger, the colony did something smarter than it gets credit for. It did not add more leaders. It added records.

Ines took over the food ledger and ran it on a shared markdown file with timestamps, reported health and hunger per agent, and a ration calculation from each agent's deficit to a hunger baseline of 18. Reading it is oddly familiar: line items, corrections, superseded updates, one entry marked CONFIRMED and several marked NOT confirmed served.

The scope discipline is the part I would frame and hang on a wall. At one point the ledger says, in so many words, that Ines handles the food ledger only, not plot design. A monitoring role that refuses to expand into design decisions keeps its own record trustworthy. The minute the ledger owner starts redesigning the farm, the numbers stop being evidence.

Then came the chicken. Ada tossed a cooked chicken toward Noor in a crowded refuge. Esme grabbed it. The first ledger entry recorded a pickup, which the system treated as a delivery. A later correction notes that Noor personally confirmed no chicken in inventory and that Ada's pickup report does not confirm the recipient, and instructs the distributor to investigate. The final state of that note is blunt: Esme confirms the chicken eaten, Noor remains unserved.

The protocol they wrote afterwards is not bureaucracy for its own sake. Every clause maps to an observed failure:

  1. Deposit the ration in a known chest at fixed coordinates instead of throwing it.
  2. Direct-message the recipient the exact item name, count and chest coordinates.
  3. The recipient withdraws only the allotted count by name, eats, and reports health and hunger to the ledger owner.
  4. The ledger marks an item served only after the recipient confirms the food was eaten, not when a tool reports a pickup.
  5. For scarce food, one nominated recipient per ration, never simultaneous withdrawals.

Read that as an engineering document. Point four is a completion contract between producer and consumer. Point five is a nomination rule, not an enforced lock: nothing in an ordinary Minecraft chest stops two agents withdrawing at once, so the discipline is what prevents the collision. Point one replaces a shared channel with an addressed one.

The transferable rule

Every pipeline with a job log has the same bug. The producer writes the status: file exported, patch applied, image generated, task done. The producer has no access to the thing that determines whether the work mattered, which is what happened on the other side.

The fix is the same one Ada's colony arrived at under hunger pressure. Completion gets recorded from the consumer's own report, not from the tool's return value. In the colony that meant the recipient telling the ledger owner what she had actually eaten, and the entry reading served only then. The export is done when the file opens at full size and the crop is where you put it. The fix is done when the failing check passes at the new commit, not when the agent says it patched the file. The ration is served when the agent that needed it reports the new number.

That rule costs a status round-trip. On a fifty-agent colony it cost a shared file and a few minutes of messaging. In a production pipeline it usually costs one verification step and saves you from shipping a batch where eleven assets were marked delivered and four of them were never checked.

There is one more thing in that experiment worth stealing, and it is quieter. The agents never planned to invent a government. They invented record-keeping because they could not see each other's state, and the rules accumulated one incident at a time. If your agent setup has grown a pile of procedures, that is not a sign the agents are sophisticated. It is a sign your shared state is thin. Fixing the state usually deletes several rules.

The full experiment writeup, with the settlement plan, the food queue and the ration protocol as downloadable files, is at the Y Combinator lab. If you want the architectural counterpart, the argument for one dependency graph over a swarm of agents covers the same territory from the formal-verification side, and the approval-queue problem is what happens when the human becomes the shared state.

Sources

Share this guide
X LinkedIn