# The Factory That Dreams: 39 AI Agents, No Framework

Rushabh Doshi

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/jtzh-GBXBWc-factory-that-dreams-39-ai-agents-no) · [Video](https://www.youtube.com/watch?v=jtzh-GBXBWc)

## The Factory That Dreams: 39 AI Agents, No Framework

Machinecraft’s Ira turns private company records into structured memory, coordinates specialist agents, and keeps a human between a drafted commercial action and its execution.

### A factory afraid of forgetting

[0:01](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=1s)

Who is this customer? What did we quote them in 2019? Why did their machine need that unusual modification? At Machinecraft, those answers were part of the company’s most valuable infrastructure, but they lived largely in people’s heads. Rushabh Doshi describes running a 100-person factory in India without a data science team or an ML budget. Its institutional memory had passed from his grandfather to his father and then to him.

Employee turnover made that arrangement increasingly fragile. Every departure took knowledge with it. The fear was less about competitors than about waking up to discover that the business depended on two increasingly tired people remembering everything.

![Slide titled “Two brains, one revolving door,” with four illustrations showing a factory, a man surrounded by papers, a revolving door, and a departing brain.](https://ai.engineer/api/talk-essay-media/fe1f7fc736db82912eef4cfdded9cc5163432e9ea9cd88dcebfde980517566c9)

Two brains, one revolving door: every departure takes company knowledge with it.

Writing another document would not necessarily solve the problem: someone still had to find it, read it, and apply it. Doshi wanted an active memory—a company twin that could use the accumulated history in everyday work. His first practical expression of that ambition was to build a sales team in software.

### The same machine enters different worlds

[1:51](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=111s)

Machinecraft makes thermoforming machines: equipment that heats a plastic sheet and shapes it. The same core machinery can produce hydroponic farm trays, spa bathtubs, EV panels, medical casings, or packaging. Doshi describes seven buyer worlds; these are the five applications he names. A useful sales system therefore needs more than brochure recall. It must recognize which industry a customer belongs to and interpret the machine’s capabilities in that customer’s context.

### Private history becomes structured memory

[2:21](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=141s)

The starting material was the company’s own history: years of quotations, drawings, payment schedules, timelines, and email threads. Doshi describes hundreds of gigabytes of private records. This was the relevant information environment—not the public web, but the records of what Machinecraft had actually promised, designed, and delivered.

There was no model training or fine-tuning. Instead, the ingestion process used off-the-shelf models to organize existing information:

1. Split the records into manageable chunks.
2. Have models read those chunks and extract facts.
3. Store representations of chunk meaning as vectors.
4. Store connections between entities as graph relationships.

The selected slide names Qdrant for vectors and Neo4j for relationships. These stores serve different purposes: semantic representations help locate relevant material, while graph connections preserve who or what is related. **The central investment is organized memory, rather than a more capable trained model.**

![Black slide headed “We never trained a model,” with a pipeline subtitle and four green illustrations of document processing, stored information, connected nodes and machinery.](https://ai.engineer/api/talk-essay-media/421cc65dc443cb47557613f9c03242a97f5383f76ab00cff60ea16b62774cd44)

“We never trained a model”: chunks, Qdrant vectors, Neo4j relationships and retrieval.

### Giving Ira a body

[3:18](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=198s)

That memory became part of Ira, a system Doshi describes as something the company was raising. Biology supplied a vocabulary for separating responsibilities:

| Biological metaphor | System responsibility |
| --- | --- |
| Senses | Identify who is interacting with the system |
| Gut | Digest documents into facts |
| Memory | Retain knowledge |
| Dream cycle | Consolidate experience |
| Immune system | Resist bad information |

The purpose of the metaphor is coherence over time. Ingestion, retention, and protection against bad information are distinct jobs, even though they all contribute to the same company memory.

### Specialists instead of one enormous prompt

[3:52](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=232s)

Doshi describes 36 agents in the spoken walkthrough. Their purpose is to divide work that would otherwise compete inside one all-purpose prompt. A pricing decision, a machine specification, and a factual correction require different responsibilities and checks.

| Agent | Responsibility |
| --- | --- |
| Athena | Coordinate the team |
| Prometheus | Own the sale |
| Plutus | Handle pricing |
| Hephaestus | Supply machine specifications |
| Vera | Fact-check |
| Memon | Guard human corrections |

For the correction guardian, Doshi sets an ambitious requirement: once a human fixes something, it should stay fixed permanently. That is the intended behavior, not a measured durability result.

The agents also collaborate. Athena convenes the relevant specialists, their contributions can disagree, and the process produces one answer. This combines narrow ownership with coordination: each agent has a bounded job, but the final response can incorporate several kinds of expertise.

### From a stranger to a customer

[5:00](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=300s)

Ira’s operating scope is the commercial front of the business: the work between discovering a potential customer and converting that prospect. Doshi reports nine daily jobs; the spoken examples identify seven:

- **Outbound email:** Draft outreach grounded in Machinecraft’s actual business context.
- **Account briefs:** Assemble cross-checked information before a call.
- **Quotations:** Prepare commercial quotes.
- **Outreach selection:** Offer a swipe-left, swipe-right review mode.
- **Lead revival:** Revisit dormant prospects through a workflow Doshi calls “blast from the past.”
- **Inbound replies:** Prepare responses to incoming messages.
- **Fit screening:** Determine whether a company is a plausible customer before spending an hour on it.

These workflows give the stored history a practical destination: information becomes preparation, qualification, or a proposed commercial action.

### One Cursor tab, with a human at the boundary

[5:45](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=345s)

The operator works from one Cursor tab. A request can trigger knowledge-base searches, inbox reading, email drafting, and quote construction. Ira then presents the work before anything goes out. The interface is small; the supporting system is not.

Behind it are separate databases for vectors, the relationship graph, and CRM records. Doshi reports using three model providers, selected according to the task. Google integrations, document ingestion, communication tools, and monitoring connect the agents to the business and make their activity inspectable.

Doshi reports 213 tools exposed through a single protocol. The governing rule is **Ira drafts; a human sends.** Tool breadth does not remove the approval boundary: preparing a message or quotation is distinct from authorizing it to leave the system.

![Slide titled “213 tools, one rule,” showing a robot filled with smaller mechanisms and the subtitle “Ira drafts. Humans send.”](https://ai.engineer/api/talk-essay-media/948271f6875333c25d6f976964a2a6136b839d4b1481231f3bb28e2034f03dd5)

213 tools, one rule: Ira drafts. Humans send.

### Remembering selectively, correcting deliberately

[6:26](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=386s)

A model’s conversational fluency does not itself provide continuity after a session closes. Ira therefore separates several kinds of memory rather than treating the current conversation as the entire company record.

| Memory layer | What it retains |
| --- | --- |
| Working memory | The last few minutes of activity |
| Pinned facts | Persistent facts about people |
| Episodes | Whole conversations as narrative records |
| Relationships | Familiarity progressing from stranger to trusted |

A **salience gate** decides what deserves retention. Without selection, persistent memory would accumulate incidental material alongside useful facts. When retained facts conflict, human corrections take precedence.

That precedence rule can be expressed independently of model behavior. In this illustrative TypeScript example, a human correction changes the payment schedule associated with a quote. A later conflicting extraction cannot overwrite it:

```typescript
type Fact = {
  key: string;
  value: string;
  origin: "extraction" | "human-correction";
};

function reconcile(current: Fact, incoming: Fact): Fact {
  if (current.key !== incoming.key) {
    throw new Error("Cannot reconcile different facts");
  }
  if (incoming.origin === "human-correction") return incoming;
  if (current.origin === "human-correction") return current;
  throw new Error("Conflicting extractions require review");
}

const extracted: Fact = {
  key: "quote-2019-014.payment-schedule",
  value: "50% deposit",
  origin: "extraction",
};

const correction: Fact = {
  key: extracted.key,
  value: "30% deposit",
  origin: "human-correction",
};

const corrected = reconcile(extracted, correction);
const retained = reconcile(corrected, extracted);
// retained.value === "30% deposit"
```

The quote identifier and deposit amounts are teaching values. The important design choice is explicit authority: correction precedence is a rule the memory system applies, rather than a request for the model to remember which answer it should prefer.

### What happens during the dream cycle

[7:28](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=448s)

Retention also has a maintenance cycle. Each night, Ira replays the day’s activity, consolidates useful information, searches for contradictions, forgets stale material, and turns work into reusable skills. In the morning, Doshi receives a dream report describing what the system consolidated, discarded, and worked out.

Doshi characterizes this as getting smarter overnight. The concrete mechanism is a scheduled process for maintaining memory and extracting reusable knowledge from experience; it is not another round of model training.

### Business principles become operating rules

[7:56](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=476s)

Each agent also has a **soul file**. Instead of relying only on generic helpfulness and harmlessness instructions, the file draws on the principles of a family business spanning three generations. Doshi describes five inherited ideas translated into engineering rules, beginning with the recognition that no single source contains the whole truth.

The resulting constraints are practical: cross-check before speaking, avoid absolute claims, cite the document and its date, stay within the assigned job, and report the truth even when it is unpleasant. Nobody works alone. Doshi describes these as production guardrails: the system’s behavior should reflect how the business wants to make decisions, not merely produce persuasive answers.

### The cost was organizing the company

[8:52](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=532s)

Doshi reports a zero model-training bill. His account of the difficult work centers on teaching the company to remember itself: extracting and organizing private knowledge, rather than training a model. He says an agency quoted “two hundred and thirty grand,” while the internal build cost was “around thirty”; he does not specify the unit for the latter figure.

Doshi reports running costs of a couple of thousand dollars per month. No itemized accounting or workload accompanies the figures, so they describe his reported experience rather than a budget another factory can directly adopt.

### Fork the architecture, supply the knowledge

[9:09](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=549s)

Machinecraft extracted the architecture into [BrainOS](https://github.com/doshirush1901/brain-os), presented as a forkable, empty nervous system. The agents, memory, dream cycle, and soul file provide the structure. An adopting company supplies its own history and knowledge; it does not receive Machinecraft’s populated memory.

That leaves the hardest responsibility with the business itself. Doshi’s closing resource is `forkmybrain.org`, but the invitation is to build a company’s own brain, not purchase his. He returns to Machinecraft’s lack of data scientists as the feasibility example: the reusable architecture can help, while the company must still decide what is true, which corrections matter, and how its accumulated knowledge should guide work.

## Resources

- [Brain OS](https://github.com/doshirush1901/brain-os) — Forkable company-brain starter with specialist agents, retrieval, CRM integration and a synthetic Acme walkthrough.
- [Ira v3 architecture](https://github.com/doshirush1901/ira-universe/blob/main/docs/ARCHITECTURE.md) — Technical map of Ira's request pipeline, specialist coordination, memory stores and biological subsystem metaphors.
- [Brain OS licensing and feature tiers](https://github.com/doshirush1901/brain-os/blob/main/docs/LICENSING.md) — Community limits, trial activation and Pro capabilities, including confirmation requirements for sending mail.
- [Ira Universe visitor demo](https://github.com/doshirush1901/ira-universe) — Public MCP package for exploring Ira's architecture and synthetic account-brief and email-draft examples.

## Slides

### IRA — Fork My Brain

Rushabh Doshi introduces how Machinecraft, a 100-person factory in India with no ML budget, built a 36-agent AI to run its go-to-market.

[0:00](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=0s) [Title slide introducing Machinecraft’s factory knowledge and AI story over a factory illustration.](https://assets.ai.engineer/sites/aiecode2025/sha256/fac703b12ca85a901922cc20fd8b55e6e5f87d5ce3efb6c0868e1ba2b3abc649.webp)

IRA · MACHINECRAFT

IRA — FORK MY BRAIN

How a 100-person factory taught itself to remember — and built a 36-agent AI that runs its go-to-market.

Rushabh Doshi · Machinecraft · AI Engineer World's Fair

### Two Brains, One Revolving Door

As employees left, Machinecraft lost pieces of its knowledge. Doshi feared the company might end up existing only in two increasingly tired heads.

[0:55](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=55s) [Four illustrations accompany the claim that the company lost knowledge when people left.](https://assets.ai.engineer/sites/aiecode2025/sha256/707d01751fdc319aff4971fed12be299027794ff14afc17df10147bd2919d9b3.webp)

02 · THE PROBLEM
TWO BRAINS, ONE REVOLVING DOOR

Every time someone left, the company forgot a little more.

### One Machine, Seven Worlds

The same thermoforming machine heats and shapes plastic sheets for products ranging from farm trays and spa bathtubs to EV panels and packaging.

[1:25](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=85s) [Illustrations of farm trays, a spa, car panels and packaging accompany a list of machine applications.](https://assets.ai.engineer/sites/aiecode2025/sha256/0c97ec0a5238caefc1beff07e36f5ab136a875022edfc72740ea5e66a7e19a42.webp)

04 · THE STAKES
ONE MACHINE, SEVEN WORLDS

Trays · spas · toilets · EV panels · medical · packaging

### Feed It Everything

The team fed the system hundreds of gigabytes of private company history, including quotes, drawings, schedules and email threads, to help it understand each customer’s world.

[2:22](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=142s) [Three illustrations accompany a list of company records used as source material.](https://assets.ai.engineer/sites/aiecode2025/sha256/ca2377046eb70bbe98d28c20dee8f9c64bc2e9b2802cfa24c17f735845c96997.webp)

05 · STEP ONE

FEED IT EVERYTHING

Years of quotes, drawings, schedules, emails. Our internet, not the internet.

### We Never Trained a Model

The team used off-the-shelf models to extract facts from chunks of company history, storing vectors in Qdrant and relationships in Neo4j for its own retrieval.

[2:40](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=160s) [Slide showing a pipeline from chunks to vectors, relationships, and retrieval, illustrated with four cartoons.](https://assets.ai.engineer/sites/aiecode2025/sha256/26459a877d03e7ccbb501e9a7b0d9c3fb144a14756f6d525c05a96ba53904757.webp)

06 · THE PLOT TWIST

WE NEVER TRAINED A MODEL

Chunk it → vectors in Qdrant → relationships in Neo4j → our own retrieval

### We Gave It a Body

Ira’s biology-inspired architecture uses senses to identify who it is talking to, a gut to turn documents into facts, and an immune system to fight bad information, alongside memory and a dream cycle.

[3:16](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=196s) [Slide listing senses, a gut, memory, a dream cycle, and an immune system above four robot illustrations.](https://assets.ai.engineer/sites/aiecode2025/sha256/4194568371515eab5ddd56322fc8f3b2d2465c957a85ea7724af17dde28dc834.webp)

07 · ARCHITECTURE
WE GAVE IT A BODY

Senses · a gut · memory · a dream cycle · an immune system

### Why Not One Genius?

Ira uses a cast of 36 specialist agents, each with one job. The speaker argues that a single prompt asked to do everything does everything badly.

[3:50](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=230s) [A lineup of illustrated characters above the statement that one prompt doing everything does everything badly.](https://assets.ai.engineer/sites/aiecode2025/sha256/a34096ea90e3500bd9630a1c3e69edd2d29895fb8b6d58850ae18281b4fb9913.webp)

08 · THE TEAM

WHY NOT ONE GENIUS?

One prompt that does everything does everything badly.

### Thirty-Six Agents, One Job Each

Ira assigns one job to each of 36 agents. Athena runs the room, Prometheus owns sales, Plutus handles pricing, and Vera fact-checks.

[4:14](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=254s) [Slide naming nine agents above a grid of nine illustrated characters.](https://assets.ai.engineer/sites/aiecode2025/sha256/64d5b6055772d6a38b835486cbbacfc9036801c45c6544a9a7f6b546ce77e840.webp)

09 · MEET THE CAST

THIRTY-SIX AGENTS, ONE JOB EACH

Athena · Prometheus · Argus · Clio · Calliope · Plutus · Hephaestus · Vera · Mnemon

### They Argue It Out

Athena brings specialists together to argue through a question and produce a single answer, pictured as a crowded boardroom.

[4:42](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=282s) [Illustrated robots and other characters argue around a boardroom table.](https://assets.ai.engineer/sites/aiecode2025/sha256/48d4b5e501e2f1722d792c4e54d5723c01ecdb7f1b3b88e836a872f355ae14e3.webp)

10 · HOW THEY DECIDE

THEY ARGUE IT OUT

A boardroom that never sleeps and has no ego.

### Nine Real Jobs, Every Day

The agents handle nine daily sales jobs, from personalized outbound emails and account briefs to reviving leads and checking whether inbound companies are a fit.

[5:11](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=311s) [Nine illustrated panels depict sales and customer workflow tasks.](https://assets.ai.engineer/sites/aiecode2025/sha256/9afaa4070f2fa694c5102e971b1cef28294778fbd18eb57e04e6f38f5c3ea030.webp)

12 · IN PRACTICE
NINE REAL JOBS, EVERY DAY

Outbound · briefs · quotes · swipe-to-send · revival · inbound · CRM · research

### One Cursor Tab

From one Cursor tab, Ira searches the knowledge base, reads the inbox, drafts emails and builds quotes, then shows the work before anything goes out.

[5:42](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=342s) [A person faces a many-armed robot in a cockpit; text says it shows work before sending.](https://assets.ai.engineer/sites/aiecode2025/sha256/2163ff1e2b5c8870896fb9cc2b2198a0ec4151fe51d1b10cef7e3151f0526dff.webp)

13 · THE COCKPIT

ONE CURSOR TAB

You ask; it reaches out with a dozen hands — and shows you before sending.

### A Real Stack, Not a Demo

Ira’s stack combines vector, graph and CRM databases with three model providers, document ingest, communications and monitoring. Each model provider is chosen for its job.

[6:03](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=363s) [Six illustrated panels accompany a list of databases, model providers, ingest, communications, and observability.](https://assets.ai.engineer/sites/aiecode2025/sha256/cac3cb6d0f3db06e2a70d5bd34da0973b3b65488c1808f340d387765f9653079.webp)

14 · THE STACK

A REAL STACK, NOT A DEMO

Vector + graph + CRM DBs · 3 model providers · ingest · comms · observability

### 213 Tools, One Rule

The system exposes 213 tools over one protocol. Ira drafts; humans send.

[6:24](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=384s) [A large robot reveals machinery inside; the caption says Ira drafts and humans send.](https://assets.ai.engineer/sites/aiecode2025/sha256/0dd82f588815df2aadf127a0dc0cd44e2669398cad119b02f9ac89f77516a8ee.webp)

15 · THE SURFACE

213 TOOLS, ONE RULE

Ira drafts. Humans send.

### Most AI Forgets You

The speaker compares a raw language model to a goldfish: brilliant briefly, then forgetting you when the tab closes.

[6:36](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=396s) [A goldfish character gestures while a robot searches a stack of drawers.](https://assets.ai.engineer/sites/aiecode2025/sha256/5c9909534862e9661f54f733a5cf0db232579b8a2bd6c030b606c935d2666531.webp)

16 · THE GOLDFISH PROBLEM

MOST AI FORGETS YOU

Brilliant for thirty seconds. Then the tab closes.

### Memory, in Layers

Memory layers hold recent context, facts, conversations and relationships. A salience gate filters what is remembered, and corrections win when facts disagree.

[6:50](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=410s) [Six illustrated panels accompany a list of working, pinned facts, semantic, episodic, relationship, and procedural memory.](https://assets.ai.engineer/sites/aiecode2025/sha256/08142bee1851bc151e865bb4e2ee0bfc6d130753e2e77a2edf7f72133eff3da1.webp)

17 · ENGINEERED MEMORY

MEMORY, IN LAYERS

Working · pinned facts · semantic · episodic · relationship · procedural

who they are
relationship warmth
open commitments

### And Then It Dreams

Each night, Ira replays the day, retains useful material, checks for contradictions, forgets stale information, and turns work into reusable skills.

[7:24](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=444s) [Illustration of small figures working beneath a crescent moon; the slide says the system gets smarter while you sleep.](https://assets.ai.engineer/sites/aiecode2025/sha256/eecdeae3856c10e97fdbffad1339eefd2d4f36ca5e0358b6e4e8c2f43c8501e6.webp)

18 · OVERNIGHT

AND THEN IT DREAMS

The thing literally gets smarter while you sleep.

### What It Does at Night

Ira’s nightly sleep cycle replays the day, consolidates useful information, prunes stale material and turns work into reusable skills. A morning dream report records what it kept, let go and figured out.

[7:31](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=451s) [Six illustrations accompany a list of overnight tasks, including replay, consolidation, pruning, learning skills and writing a dream report.](https://assets.ai.engineer/sites/aiecode2025/sha256/5ef497d0cd52cce7497a47b4d968d2ed30452229ee1614dbdee2f64663b28a23.webp)

19 · THE NIGHT SHIFT
WHAT IT DOES AT NIGHT

Replay · consolidate · prune · learn skills · write the dream report · wake up

### Five Old Ideas, Running in Prod

Each agent has a soul file that turns five principles into guardrails: cross-check sources, qualify and cite claims, do its own job, report uncomfortable truths, and work with others.

[8:18](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=498s) [Five illustrations accompany principles including qualifying claims, reporting truth and serving each other.](https://assets.ai.engineer/sites/aiecode2025/sha256/516141de826a30cf94728135db528efc9d3c587ec110a4fc657a872edd0c18d0.webp)

21 · THE GUARDRAILS

FIVE OLD IDEAS, RUNNING IN PROD

Many-sidedness · qualify claims · do your duty · truth over optics · serve each other

### Fork an Empty Brain

BrainOS makes the architecture forkable: its agents, memory, dream cycle and soul file start blank, ready for a company to fill with its own knowledge.

[9:08](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=548s) [Three illustrations accompany the statement that Brain OS ships blank for users to fill with their own knowledge.](https://assets.ai.engineer/sites/aiecode2025/sha256/040679b7788e4eef3332c5df2e09dacc05a757aeae2e7c8179f206ff62c5b75a.webp)

23 · THE MOVE

FORK AN EMPTY BRAIN

Brain OS ships blank. You pour your own truth in.

### Build Your Own Brain

The speaker points to a 100-person factory with no data scientists and invites others to build their own brain.

[9:35](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=575s) [Bright closing slide invites viewers to build something that remembers.](https://assets.ai.engineer/sites/aiecode2025/sha256/2e85851073bcfe54b2e493fa9449e82b38d0945e19e3903b22f0c01a337046d2.webp)

forkmybrain.org

We don't sell you our brain. We help you build your own.

Go build something that remembers.

## Full transcript

[0:01](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=1s) Okay, I want to tell you a story about a factory that taught itself how to remember. Hi, I'm Rushabh. I run Machinecraft, a hundred people factory in India. No data science team, no ML budget, none of that.

[0:15](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=15s) And somehow, we ended up building a 36 AI agent that runs our entire go-to-market. I think that's still a rid-- little ridiculous. Let me show you how it happened and why you can do the same thing.

[0:30](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=30s) So here's the thing about our company. From the outside, it looks like machines and metal. But the actual company, the part that matters, isn't the machines. It's the knowledge.

[0:40](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=40s) Who the customer is, what we quoted them in 2019, why that one machine needed that weird custom tweak. And for three generations, all of that lived in exactly three brains.

[0:51](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=51s) Initially my grandfather's, then my father's, and now mine.

[0:57](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=57s) Which is a genuinely terrifying way to run a company when you sit with it.

[1:02](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=62s) A lot of people have joined us, people have left us. The revolving door never stopped.

[1:08](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=68s) And every single time someone walked out, a chunk of our brain walked out with them.

[1:15](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=75s) We weren't scared of the competitors. We were scared of forgetting or waking up one day and realizing the whole company only existed inside two increasingly tired heads.

[1:27](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=87s) So I had an idea. I'll be honest, it sounded insane first. But what if instead of writing the knowledge down in some document nobody ever reads, what if we grew a brain that just held it?

[1:42](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=102s) Not a chatbot you poke at, a twin of the company.

[1:46](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=106s) I didn't hire a sales team. I tried to build one.

[1:51](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=111s) A quick detour, because you need to know how messy this is. We make thermoforming machines. They heat up a plastic sheet and shape it. Same core machine, but it ends up making hydroponic farm trays, spa bathtubs, EV car panels, medical casings, and even packaging.

[2:10](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=130s) Seven totally different worlds, seven totally different buyers. So this brain couldn't just memorize a brochure. It had to know which universe

[2:21](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=141s) a given customer lives in. Step one was almost boringly simple. Feed it everything, and I mean everything. Years of quotes, drawings, payment schedules, timelines, email threads, hundreds of gigabytes of our own private history.

[2:38](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=158s) Not the public internet, our internet. And here's the plot twist, the part that surprises every engineer I tell this to. We never trained a model. No GPUs humming in the basement, no fine-tuning.

[2:53](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=173s) We just looked at all the history, chopped it into bite-sized chunks, and let off-the-shelf models read it and pull out the facts. We stored the meaning of each chunk as vectors and relationships.

[3:07](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=187s) Who's connected to what as a graph? The brain isn't a smarter model. It's actually a really, really well-organized memory.

[3:18](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=198s) Now, this is where it gets a little weird in a good way. We stopped thinking of Ira as a software and started thinking of it as something we were raising.

[3:27](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=207s) So we gave it a body modeled on biology. Senses to figure out who it's talking to, a gut to digest the documents into facts, a memory, a dream cycle, an immune system to fight off bad information.

[3:42](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=222s) Why biology? Well, because evolution already spent a billion years solving how do you stay coherent over time. We just copied the homework.

[3:52](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=232s) Okay, so the big question, why 36 agents instead of one genius mega prompt? Because, and you already know this if you've ever tried it, one prompt that's supposed to do everything ends up doing everything badly.

[4:07](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=247s) So Ira isn't one mind, it's a pantheon, a whole cast of specialists. Each one has exactly one job. Athena runs the room. Prometheus owns the sale. Plutus does pricing.

[4:23](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=263s) Hephaestus knows every machine spec cold. Vera fact checks everything. And Memon, my favorite, guards corrections, so the second a human fixes something, it stays fixed forever. One agent, one job.

[4:40](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=280s) It's a team, not a hero. And here's the cool part. They hold meetings. Athena pulls in specialists. They actually argue, and a single answer comes out the other side.

[4:54](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=294s) It's like having a boardroom that never sleeps, never gets tired, and somehow has no ego.

[5:00](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=300s) So what does all this actually run? Honestly, the whole front business. Everything between a stranger exists somewhere and now they're a customer.

[5:13](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=313s) Nine concrete jobs every single day. Outbound emails that actually reference my real world. Account briefs built from cross-checked truths before a call. Quotations. A swipe left, swipe right mode for outreach.

[5:29](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=329s) Reviving dead leads, which I call ba- blast from the past. Inbound replies, and figuring out before we waste an hour, whether a company is even a fit. Nine jobs, one operator who never sleeps.

[5:45](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=345s) Where does all this live? One Cursor tab. That's genuinely it. You type and Ira reaches out with a dozen hands, searches the knowledge base, reads the inbox, drafts the email, builds the quote, and then shows you before anything actually goes out.

[6:03](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=363s) Under the hood, it's genuinely a real stack, not a demo held together with a tape. Databases for vectors, for relationship graph, for the CRM. Three different model providers, each picked for the job it's actually best for.

[6:17](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=377s) Tools for Google, for swallowing documents for every communication channel, plus monitoring, so we can see what it's thinking.

[6:26](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=386s) All of it, every capability exposed as two hundred and thirteen tools over one protocol. And the golden rule, the one we never break, Ira drafts; human sends it. Now, memory, and this is the part where most AI quietly lies to you, because a raw language model is basically a goldfish, brilliant for about thirty seconds, and then you close

[6:49](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=409s) the tab and forgets you ever existed. So we engineered memory on purpose in layers, working memory for the last few minutes, pinned facts about someone who, whiz. Episodes, whole conversations as little stories, relationships with warmth that grows from stranger to trusted, and a bouncer at the door, a salience gate that

[7:14](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=434s) decides what's even worth remembering so the brain doesn't fill up with junk. When two facts disagree, corrections win. Continuity without making things up. And then, I genuinely love this part.

[7:28](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=448s) At night, it dreams. Every night, Ira runs a sleep cycle. It replays the day, locks in useful stuff, hunts for contradictions, gently forgets the stale junk, and turns the day's work into reusable skills.

[7:45](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=465s) In the morning, there's a little dream report waiting for me to read. Here's what I consolidated. Here's what I, here's what I let go of. Here's what I figured out while you were asleep.

[7:56](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=476s) The thing literally gets smarter overnight. And here's the part I care about the most. Every agent has a conscience, and it is emphatically not to be helpful, be harmless.

[8:12](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=492s) It's a soul file written from the principles of a [REDACTED:religion] family business that's been doing this for the last three generations. Five old ideas turned into engineering rules. No single source has the whole truth.

[8:25](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=505s) So cross-check before you speak. Never state things absolutely. Cite the document and the date. Do your own job, not someone else's. Report the truth, even when the truth is ugly.

[8:38](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=518s) And nobody works alone. Ancient philosophy running as guardrails in production. Now let's talk money, because this is the part that should make the whole industry a little uncomfortable. There was no training bill.

[8:52](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=532s) Zero. The expensive part was never compute. It was teaching a company to remember itself. An agency quoted us two hundred and thirty grand to build this. We built it for around thirty.

[9:03](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=543s) That's cheaper than a nice watch. And it runs on a couple of thousand dollars a month.

[9:09](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=549s) So here's the move. We pulled the whole architecture out and made it forkable. We call it BrainOS. It ships as an empty nervous system. The agents, the memory, the dream cycle, the soul file, all there, completely blank.

[9:24](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=564s) You pour your own company's truth into it and it grows from the inside out. Because here's the thing nobody can outsource for you. Only you can build your company's brain.

[9:37](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=577s) We are a hundred people factory with no data scientists. If we can grow a brain, you can too. We're not selling ours to you. We're helping you build your own, forkmybrain.org.

[9:49](https://www.youtube.com/watch?v=jtzh-GBXBWc&t=589s) Go build something that remembers. Thank you.
