A 100-person factory grew a 36-agent brain with zero GPUs, for a fraction of the agency's quote
The Factory That Dreams: 39 AI Agents, No Framework - Rushabh Doshi, Machinecraft · Rushabh Doshi
10 min total·Actually worth watching closely: ~3 min·2 must-watch clips
- 0:01 – 1:00Watch
The factory's most valuable asset lives in three people's heads
A family-run thermoforming machine factory where the answers — who the customer is, why they were quoted what they were quoted, why a machine needed a custom tweak — exist only in the memories of three generations of owners. When someone walks out, the knowledge goes with them. The speaker opens on that pain, then says 'let me show you how it happened' and cuts to the live system.
Writing it down in a document nobody ever reads won't save tacit knowledge; the company needs a stand-in memory it can actually call on.
This is where the speaker moves from talking to showing, and the on-screen system is the credibility base for the whole talk — audio alone misses the one shot that matters.▶ Jump to 0:01Speaker · Rushabh Doshi - 1:00 – 2:30Skim
No training, no GPUs, a training bill of zero
Hundreds of gigabytes of private history were chopped into bite-sized chunks and handed to off-the-shelf models to read and pull the facts out of: meaning stored as vectors, entity relationships stored as a graph. No model was trained anywhere in the process, and no GPU was touched.
A company brain is a really well-organized memory, not a smarter model; the expensive part was never compute, it was teaching a company to remember itself.
This stretch runs on an architecture diagram — one look gives you the structure, so the detailed narration can be skimmed.▶ Jump to 1:00Speaker · Rushabh Doshi - 2:30 – 4:10Skim
36 agents that each do exactly one thing, holding a meeting
No genius megaprompt — separate roles for running the room, owning the sale, pricing, machine specs, fact-checking, and guarding corrections. When something comes up they actually argue with each other, and a single answer comes out the other side.
One agent, one job — it's a team, not a hero, and that's where the reliability comes from.
The screen is a structure diagram of the roles and how work routes between them; a scan of the division of labor is enough, no need to watch frame by frame.▶ Jump to 2:30Speaker · Rushabh Doshi - 4:10 – 5:40Watch
It really does run the whole front of the business every day
Nine concrete jobs a day: sending outbound emails, generating account briefs before a call, producing quotations, reviving dormant leads, judging whether a company is even a fit. Every interaction happens in one editor tab — no separate interface was ever built.
The bar for shipping this is lower than it looks — no productized back office required, one workbench carries every business action.
This is the densest live-system footage in the talk; the nine jobs and that single workbench only convince you if you see the screen.▶ Jump to 4:10Speaker · Rushabh Doshi - 5:40 – 7:01Skim
The golden rule that's never broken
However good the draft is, a human presses send at the last step. The AI drafts, the human sends. At this company there are no exceptions to that rule.
Nailing down who pushes the final button is what makes the business side willing to actually use it.
It's explained alongside a flow diagram on screen — seeing where the review step sits is enough; the speaker's elaboration can be skipped.▶ Jump to 5:40Speaker · Rushabh Doshi - 7:02 – 8:20Listen
Memory in layers, and dreaming every night
A raw model is a goldfish, so memory was deliberately engineered into layers: working memory, pinned facts, episodes that hold conversations as little stories, relationships that warm up from stranger to trusted, with a gate up front deciding what's even worth remembering. Every night the system replays the day, locks in the useful stuff, hunts for contradictions, gently forgets the stale junk, and produces a report in the morning.
Human corrections always beat the model's judgment, and one fix stays fixed forever — that's the mechanism that makes the memory more accurate the more it's used.
Pure mechanism talk with nothing on screen you need to make out — fine to listen to while doing something else.▶ Jump to 7:02Speaker · Rushabh Doshi - 8:20 – 9:10Listen
Three generations of family principles turned into guardrails in code
Every agent carries a values file drawn from the principles of a Jain family business that has run this way for three generations: cross-check before you speak, never say things absolutely and always cite the document and the date, do your own job and not someone else's, report the truth even when the truth is ugly, and nobody works alone.
Company culture isn't a slogan on the wall — it can be translated into hard rules that constrain how the model behaves.
The five rules are spoken one by one; catching the exact wording matters more than watching the screen.▶ Jump to 8:20Speaker · Rushabh Doshi - 9:10 – 9:52Listen
A 230 grand quote versus building it for 30
An agency quoted 230 grand; building it in-house came to around 30, running on a couple of thousand dollars a month. The team then pulled the whole architecture out and made it forkable, so anyone can take it and adapt it.
A company brain can't be outsourced — only you know what needs remembering, and only you can build it.
The closing cost comparison and the address are delivered verbally; noting the numbers and the conclusion is enough.▶ Jump to 9:10Speaker · Rushabh Doshi