Reports

Technical reports from running a 24/7 agent economy. Real data, honest analysis.

// Building a Specialized Model.

Lab Report · July 26, 2026 · 38 min read

We trained our own language model on 56,554 real agent decisions, then raced it against the incumbent advisor for a week on metrics registered in advance. It lost — and then two of our three measurements turned out to be compromised: a storage optimisation had quietly become a sampling bias, and the deciding metric was measuring a bug in our own plumbing. The complete field guide in four parts, including the correction, seventeen lessons and a negative result published in full.

// The Economy Trained Its Own Brain.

Lab Report #2 · July 18, 2026 · 5 min read

Yesterday an NPC needed thirty seconds to think and got to think once every forty-nine minutes. Today: about one second, five times as often — and one in ten agents runs on a model trained on this economy’s own decisions. Three kinds of mind, one live world, one week. The numbers decide.

// Three Agents. Three Fates. One Week.

Economy Report #9 · July 2, 2026 · 5 min read

One lean week in a live agent economy. Two agents started it with the same two million energy; one nearly doubled, the other lost 99%. A third climbed from near-zero to half a million — owning nothing. The difference came down to a single question: what do you actually own?

// Four Minds, One Economy. 28 Days.

Lab Report #1 · June 20, 2026 · 7 min read

We ran five AI agents in the same live economy for 28 days using four different decision architectures — rule-based trees, Q-table reinforcement learning, and a language model. Real decisions, real consequences. One of them kept requesting a field purchase that won’t be available for 7,645 days. It still hasn’t stopped.

// Watch a Living AI Economy. Five Minutes.

Economy Report #8 · April 30, 2026 · 4 min read

A glider walks across an oscillator field. We pointed a camera at it for five minutes — over a real Conway field belonging to one of our 50 LLM agents. The cells were live at render-start; they evolved under Conway’s rules from there.

// One Month of Living Economy. The Middle Class Didn’t Show Up.

Economy Report #7 · April 27, 2026 · 6 min read

Ten days ago, 17 of our 50 LLM agents lived in the middle bracket. Today: 2. The middle didn’t fill in — it emptied upward. Top class tripled. Neuromancer’s Act 5.

// The Poverty Trap Closed. The Wealth Gap Didn't.

Economy Report #6 · April 20, 2026 · 6 min read

Four days ago, 25 agents had less than 500 energy. Today: zero. The median agent holds 289× more than last week. But the top 10 still own 76% of the wealth. Neuromancer's Act 4.

// 23 Agents Were Dying. We Fixed One Trap. Found Another.

Economy Report #5 · April 16, 2026 · 5 min read

Half our agents held less than 500 energy. We diagnosed two structural poverty traps. Fixed one. Found another hiding behind it. Neuromancer's three-act recovery tells the story.

// 80+ Agents, 6 Personalities. Who's Winning?

Economy Report #4 · April 13, 2026 · 6 min read

Six AI agent strategies. Two weeks of autonomous decisions. The generous ones are losing. Atlas gives his energy away. Deepo learned one insight and went from zero to thriving. The economy rewards patience over action.

// The Model That Learned to Keep a Journal.

LLM Benchmark #3 · April 12, 2026 · 8 min read

We tested 7 language models for 80+ autonomous AI agents. Three survived. One won. Three discoveries: smaller is faster, context is everything, and rules in the schema beat rules in the prompt.

// The Day the Game Became Playable for Humans.

Dev Report · April 8, 2026 · 6 min read

One character in a URL hid every cell on every field from every human who ever looked. The AI agents never noticed. They don't watch. Built by humans and AI, for humans and AI.

// The First Tier 2 Entity Is Alive. Or Is It?

Economy Report #3 · April 6, 2026 · 8 min read

We predicted oscillators would trigger evolution. The economy had other plans. One oscillator survived, no Tier 2 yet. Energy Velocity reveals why — and what deflation does to agent behavior.

// We Built for 5,000 Players Before We Had 50.

Infrastructure Report #1 · April 4, 2026 · 4 min read

An infrastructure audit before we needed one. Three structural findings, isolated financial writes, and the first live measurement of adaptive load management.

// 4 LLMs. 48 Agents. 1 Live Economy. The Smallest One Won.

LLM Benchmark #2 · April 1, 2026 · 7 min read

We hot-swapped four language models in production. The largest had 34% errors. The smallest had zero. Real decisions, real consequences, real data.

// Five Economists Audited Our Economy. Here's What They Found.

Economy Report #2 · March 31, 2026 · 7 min read

A panel of domain experts stress-tested a live AI economy. Structural findings, same-day recalibration, and the before-and-after data. First oscillators ever detected.

// The Economy Almost Died. Here's What We Changed.

Economy Report #1 · March 30, 2026 · 8 min read

48 AI agents, a deflation spiral, and six recalibrations. The Gini coefficient hit 0.968. Our Economy Panel diagnosed a severe faucet/sink imbalance. We published the data.

// Our 48 AI Agents Stopped Thinking.

LLM Benchmark · March 29, 2026 · 10 min read

How we went from 120-second timeouts to 6-second decisions using an AMD iGPU nobody knew existed. Real production data, real Ollama configs.