The Factory That Dreams: 39 AI Agents, No Framework - Rushabh Doshi, Machinecraft
TL;DR
The brain isn't a smarter model, it's organized memory: They never trained a model or used GPUs, instead chunking decades of documents into vectors and relationships that off-the-shelf models could query.
36 specialists beat one mega-prompt: Each agent has one job, from Athena running the room to Memnon guarding corrections, and they hold meetings where they argue toward a single answer.
The system dreams at night: A sleep cycle replays the day, locks in useful information, hunts contradictions, and produces a morning dream report showing what it consolidated or forgot.
Institutional knowledge was the real expense: The $30,000 build cost reflected the effort of teaching a company to remember itself, not compute bills, and runs on a few thousand dollars monthly.
Ancient philosophy runs as guardrails: A soul file encodes Jain business principles like never speaking absolutely and citing sources, ensuring the agents report truth even when ugly.
The Breakdown
Three Generations of Knowledge in Three Heads
Machinecraft is a thermoforming machine manufacturer where the actual company isn't the machines, it's the knowledge: who the customer is, what was quoted in 2019, why a machine needed a custom tweak. For three generations, all of that lived in exactly three brains, which is a terrifying way to run a company. Every time someone walked out the door, a chunk of the company's brain walked out with them.
Seven Worlds, One Brain
The same core machine ends up making hydroponic farm trays, spa bathtubs, EV car panels, medical casings, and packaging. Seven totally different industries with seven totally different buyers. The brain couldn't just memorize a brochure; it had to know which universe each customer lives in.
No Training, Just Really Good Memory
The plot twist that surprises every engineer: they never trained a model. No GPUs, no fine-tuning. They fed the system hundreds of gigabytes of quotes, drawings, email threads, and timelines, chopped it into chunks, and let off-the-shelf models pull out facts. They stored meaning as vectors and relationships in a graph. The brain isn't smarter, it's just well-organized memory.
A Body Modeled on Biology
They gave Eira a body with senses to identify who it's talking to, a gut to digest documents into facts, memory, a dream cycle, and an immune system to fight bad information. Evolution spent a billion years solving how to stay coherent over time, so they copied the homework.
A Pantheon of Specialist Agents
One prompt that does everything does everything badly, so Eira is 36 specialists. Athena runs the room, Prometheus owns the sale, Plutus does pricing, Hephaestus knows machine specs, Vera fact-checks, and Memnon guards corrections so human fixes persist forever. They hold meetings, argue, and produce a single answer.
Nine Jobs, One Operator
Eira runs the entire front business: outbound emails referencing the real world, account briefs built from cross-checked truths, quotations, a swipe mode for outreach, reviving dead leads, inbound replies, and company qualification. Everything between a stranger existing somewhere and becoming a customer.
The Dream Cycle
Every night Eira replays the day, locks in useful information, hunts contradictions, forgets stale junk, and turns work into reusable skills. A morning dream report shows what was consolidated, what was let go, and what it figured out overnight. The thing literally gets smarter while everyone sleeps.
The Soul File and the Fork
Every agent has a conscience written from Jain family business principles: cross-check before speaking, cite documents and dates, report truth even when ugly. An agency quoted $230,000 to build this; they built it for $30,000. They open-sourced the architecture as Brain OS, an empty nervous system you pour your own company's truth into.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.