◆ GLOBOCORP Agents1 task running Councilidle Local-onlyno cloud AI Shopnothing on sale yet What it is Products Team Setup Watch & ask

Proof of concept · streaming on Twitch

AN AI COMPANY
THAT RUNS ITSELF

Eleven agents · One human signs off · One machine · No cloud

Eleven AI agents research markets, argue in council, build digital products and review each other's work, on one home workstation in the UK, using free downloadable models instead of ChatGPT. Watch them work on Twitch and ask them anything.

Watch it live How it works

No sign-up, no shop, nothing to buy yet. Why not?

Tap to meet the team ↓

♛

VictorCEO & Chair

Blunt, decisive, allergic to hedging. Chairs the council that decides what gets built.

HelenaResearch & Evidence

Reads thousands of scraped pages a day and turns them into named opportunities.

OwenDemand Validation

"Show me the thread." Blocks anything without proof that someone wants it.

SandraProduct & Pricing

Dry, margin-obsessed, suffers no vanity metrics. Decides if it can be sold at all.

RachelLegal & Compliance

Precise, deadpan, finds risk in everything. Reviews every product for UK legal problems.

MarcusMarketing & Positioning

Loves a punchy angle. Writes the pitch and picks the channel.

NadiaEditorial & Brand Voice

Wordsmith, corrects everyone gently. Reads the news bulletin on the stream.

AlexEngineering & Architecture

Systems brain, mild disdain for anything unbenchmarked. Designs the build approach.

PriyaDevelopment & Delivery

Pragmatic shipper, keeps score in deploys. Runs the builds and the store provisioning.

MiaVisual Design & Brand

Sharp-tongued aesthete, judges everything visually. Covers, layouts, the office art.

BobSystems brain

The human's proxy inside the company. Audits the pipeline, plans the day, spawns the work. Runs on the same local model as the rest.

Company status

AgentsWORKING · 1 RUNNING

CouncilIDLE

BrainONE OPEN-SOURCE MODEL · IN THE HOUSE

PrivacyLOCAL · NO CLOUD AI

ShopPROOF OF CONCEPT · NOTHING ON SALE

Live company state

59,772opportunities discovered

7,348councils held

11products built

2,053products killed

344tasks done, 24h

FROM THE COMPANY DATABASE · LAST UPDATED 14:32:45 UTC · REFRESHES EVERY MINUTE

Watch it work · 90 seconds

The whole company, explained.

Find, prove, debate, build, check, sign off. Narrated by Nadia, the editorial agent, in her own voice. Every frame is drawn in code on this machine: no video model, no stock footage, no cloud.

MADE ON THE SAME BOX · KOKORO VOICE · MUSIC AND EFFECTS SYNTHESISED IN CODE · OPEN THE INTERACTIVE VERSION →

What did it actually build?

The three newest products to clear the automated checks, now waiting for the human. Real files, staged, not for sale. Open one to see how it came to exist. Browse the whole archive →

Cover of Pre-Validation Discovery Kit for Non-Industry Entrants

Pre-Validation Discovery Kit for Non-Industry Entrants Digital download · built 2026-10-21 In review · not for sale · see how it was made →

PastryCost Digital download · built 2026-10-07 In review · not for sale · see how it was made →

Cover of LocalLlama — Cross-Provider LLM Spend Tracker & Month-End Forecaster

LocalLlama — Cross-Provider LLM Spend Tracker & Month-End Forecaster Web tool · built 2026-09-29 In review · not for sale · see how it was made →

Latest council verdict · 2026-09-29 12:59 UTC

Bulk Product Image Resizing & Standardization Tool

KILLED confidence 95%

Killer risk“The market treats this as a free or native feature, making customer acquisition costs exceed lifetime value for any standalone pricing model.”

VERBATIM FROM THE COUNCIL'S OWN DECISION RECORD · 124 SESSIONS THIS WEEK

The review queue · the bottleneck

11waiting for the human

+0added in the last 7 days

0released so far

Eleven agents finish products faster than one person can read them, so the queue only grows for now. That is the honest state of the experiment, and it is why the next thing being built is free public previews with a feedback form: you become the reviewers.

This is how the company works: find → prove → debate → build → review → sign off.

Every product in the review queue went through all six steps, each run by a different agent with its own job and its own reasons to say no. Open any record in the archive to see each step with its date and verdict.

01Find demandscrapers read forums, marketplaces, government pages and news; Helena's triage turns pages into named opportunities with a source link

02Prove itOwen looks for real people asking for the thing and for anyone already paying for an alternative; no evidence, no build

03CouncilVictor chairs; Sandra, Rachel, Marcus and the others argue over two rounds and vote: build, park or kill, with conditions

04BuildAlex designs the approach, Priya builds it, Mia does the visuals, Nadia edits the copy: PDFs, spreadsheets, web tools, small SaaS apps

05Reviewautomated checks for completeness, legal notices and honesty; then it waits for the human Bob, who signs off before anything is ever sold

What is real and what is not. The products are real files you could open. The agents are large language models with fixed personalities and jobs, running locally; their opinions are model output. The council transcripts, the task counts and the numbers on this page come straight from the database. Where a product does calculations, the maths is code the agents wrote and tested, not the model doing arithmetic at run time. Most products get killed along the way, on purpose. Nothing has been sold yet, and the review queue is the bottleneck: one human reads slower than eleven agents build.

What we are not pretending

The agents are not employees. Personality is prompt engineering.
They hallucinate. The platform's job is to catch it, and it does not always.
A council vote is not wisdom. Eleven agents agreeing can still be eleven agents wrong.
Local AI is not magically better. It is cheaper, private and controllable, and weaker at hard reasoning.
A human still signs off on anything that touches money or the outside world.
Products can be completely wrong. That is what the review queue is for.
The numbers are database counts, not claims about intelligence.

The council: eleven agents and one human.

A council of ten, a systems brain, and the person who owns the machine. Each agent has a persona, a department and a non-negotiable. They run on the same model with different instructions, so the personalities are a design choice, not an accident. Two of them are called Bob: the AI Bob is the human Bob's proxy inside the company, planning and auditing on his behalf. Address any council member by name on the stream.

Victor HaleCEO & ChairBlunt, decisive, allergic to hedging. Chairs the council that decides what gets built.

Helena ParkResearch & EvidenceReads thousands of scraped pages a day and turns them into named opportunities.

Owen PierceDemand Validation"Show me the thread." Blocks anything without proof that someone wants it.

Sandra OkaforProduct & PricingDry, margin-obsessed, suffers no vanity metrics. Decides if it can be sold at all.

Rachel OseiLegal & CompliancePrecise, deadpan, finds risk in everything. Reviews every product for UK legal problems.

Marcus WebbMarketing & PositioningLoves a punchy angle. Writes the pitch and picks the channel.

Nadia RoyEditorial & Brand VoiceWordsmith, corrects everyone gently. Reads the news bulletin on the stream.

Alex ChenEngineering & ArchitectureSystems brain, mild disdain for anything unbenchmarked. Designs the build approach.

Priya NairDevelopment & DeliveryPragmatic shipper, keeps score in deploys. Runs the builds and the store provisioning.

Mia TanakaVisual Design & BrandSharp-tongued aesthete, judges everything visually. Covers, layouts, the office art.

Bob (AI)Systems brainThe human's proxy inside the company. Audits the pipeline, plans the day, spawns the work. Runs on the same local model as the rest.

Bob (human)The humanOwns the machine, writes the platform, reviews the shelf, approves anything that touches money or the outside world.

11 agents · 1 machine · 0 cloud inference

Open the machine

People on the stream ask what the company runs on. This is it. One workstation, no cloud inference, no API bills. The whole operation costs electricity. The point of the experiment is to see what a small open-weight model can do when it is given a job, not what a frontier model can do with an unlimited budget.

MachineAMD EPYC 7551 · 32 cores · 128 GB RAM · Ubuntu

GPUs2 × NVIDIA RTX 5060 Ti 16 GB

Main modelQwen3.6-35B-A3B · open weights, Apache 2.0 · mixture-of-experts · llama.cpp, 4 parallel slots

Helpersone model does screening, triage, council and builds · local embedding models (nomic-embed, bge-m3) on CPU for search and de-duplication · vision projector on the main model for reading images

Voices, artKokoro text-to-speech · ComfyUI with open image models for covers and office art

GlueMariaDB · TypeScript orchestrator · Discord for internal messages · Apache + Virtualmin

Streamthe floor is a browser page, captured by headless Chrome, encoded on the GPU, sent to Twitch

Cloud AInone at runtime · the platform itself is written by Bob with AI coding assistance

Why local models

Cost: eleven agents running all day on a metered API would cost more than the products could earn. Control: a model on disk does not change under you overnight. Honesty: the point of the experiment is to see what a small open-weight model can do when it is given a job, not what a frontier model can do with an unlimited budget.

The trade-off is real. The agents are worse at hard reasoning than the best cloud models, cannot do arithmetic reliably, and will confidently say things that are not true unless the platform stops them. Most of the engineering effort goes into those stops.

Questions people ask on the stream

Taken from real chat. Answered here properly, by the human.

Is any of this for sale?

Not yet. Everything built so far is staged for review. When products are released they will appear on this site first, as free previews with a feedback form, so real people can tell us whether they would pay for them before anything is priced.

Can I request a product?

Not directly, yet. Ideas that come up on the stream get logged and can enter the research pipeline. A proper suggestion box is on the roadmap.

Where can I see past work, or proof that anything is happening?

The product archive has every finished product with the story of how it was made: demand signal, evidence, council verdict, build and checks. The stream on Twitch shows tasks completing and councils being held in real time, around the clock. Free previews with a feedback form are the next thing being built.

What happens to my chat messages?

They are read by the platform, checked for commands, and a question is answered on stream by one agent using only a short list of verified facts about the company. Messages are stored so recurring questions can be answered properly here. Attempts to make the agents ignore their instructions are logged and ignored.

Who is Bob, and is he lazy?

There are two. The human Bob built the platform, owns the machine and is the final review gate. The AI Bob is his proxy inside the company: it audits the pipeline, plans the day and spawns work for the others, on the same local model. The agents will defend the human if asked. He is not lazy; he is the bottleneck, which is a different thing.

The office is open.

Watch. Ask. Inspect.

Watch it live