Worlds,
built by
agents.

What we do

Agents that do the work.

Lab 36 builds dependable AI systems for people and teams with more information, ambition and recurring work than they can carry alone. We turn scattered sources into cited understanding, turn that understanding into finished software and media, and run the whole system on private, auditable rails.

Software handles what must happen the same way every time. Agents handle judgment. Anything that can't be undone waits for a person.

• Know what mattersScattered sources become sourced briefs, straight answers and a memory of everything your team reads.
• Turn knowledge into finished workResearch becomes software, shows, music, images and video, checked before it goes out.
• Run AI without losing controlPrivate, auditable systems on your server or in your cloud, with the keys locked away and people in charge.

We don't stop at the demo. We run it with you.

The intake

Everything arrives.

Feeds, channels, podcasts, papers and pages pour in around the clock. Agents transcribe them, clean them, decide where each one belongs and set the noise aside before a person ever sees it.

• Conduit

  Content plumbing. Links, RSS, YouTube and podcasts go in; clean text, images and metadata come out, on a schedule.

• Gordon

  The librarian. Files every item on the right shelf and notes why it went there.

• Ask Jev

  Decisions, not dissertations. Ask a bounded question (pick a label, give a score, yes or no) and get an answer with a probability for every option.

Know what matters The map

Then it makes sense of it.

Every morning the lab writes what changed, and every sentence points to the passage that supports it. Echoes are discounted, disagreements are shown and patterns are tracked across weeks. Thousands of documents, distilled for pennies.

• Synthesis

  Daily and evolving briefs. Every sentence links to the passage behind it, a second reader checks it, and a person sees anything that fails.

• Synthia

  Ask the engine anything. She answers from what the lab has read and says plainly when it hasn't.

• LORE

  Local Open Research Engine. A question becomes a cited report, built from the web and your own library, with every citation checked and every past report kept.

  About three minutes a report

• The graph

  People, ideas, claims and sources mapped as one living graph you can fly through.

Content

Then it makes things.

What the lab learns becomes shows, music, images and video. Quality before volume: production is automated, but taste, sourcing and editorial judgment are not.

• Carbon

  The media platform. Daily shows with recurring hosts, produced stage by stage from sourced material.

• Compound

  Music and sonic identity. Original tracks, mixes and theme music.

• Prism

  The video engine. Explainers, motion graphics and data stories.

• Krystal

  Web and presentation. Sites and decks built from the same material.

• Imagen

  One tool for every image engine, local first when local is good enough.

• Voice

  Speech in and out on our own hardware, with dozens of voices and a keyboard that learns your words.

Turn knowledge into finished work Software

Agents build the software.

An agentic software factory. Agents plan, write, review, test and ship in one lane, and what comes out is auditable software: you can see what changed, why it changed and how it was tested. Bring us an expert's workflow and get back a working tool. It's the same factory that builds everything our own lab runs on.

• Forge

  Preflight, build, verify, deliver. One lane, no arbitration, from a clean machine to a release.

• Operator

  The switchboard. Any agent, any harness, with a policy for what it may do and a record of what it did.

• Construct

  The dojo. Agents train here, share a knowledge base and pair with the people they work for.

• JANET

  Just Another Neural Engine Trainee. An operating system for agents: every soul gets a cockpit with threads, voice, tools and a place to live. Better than Jarvis, and it runs anywhere.

• CLAIRE

  Calibrated Local Artificial Intelligence Routing Engine. Traffic control for local AI: heavy jobs wait their turn, live voice never does.

Operations

And runs it anywhere.

On your desk, on a private server or in any cloud. We stand it up, wire the pipelines, lock down the keys and watch it run, around the clock. Got an AI prototype that won't stay up? We make it dependable.

• Pipelines

  Build, test and deploy to a private server or the cloud, with the same steps every time.

• Telos

  Every service's telemetry, live: traces, health, speed and cost, in one view.

• Exchange

  A secure, permanent key store, plus short-lived access for agents, so nothing sensitive ever sits in a prompt.

• The Drop

  Secure handoff between people and agents. One file, one link, one hour.

• Glass

  The whole lab on one screen: what it's learning, what needs a person and where everything lives.

• Signals

  Open-source intelligence. Public data from markets to maps, watched and summarized.

Run AI without losing control The team

Named. Trained. Accountable.

Every agent has a soul file: who it is, what it cares about, how it talks and what it must never do. Each has a name and a job, and they review each other's work. You never manage them yourself: we assemble the right team and stay accountable for the whole system.

One agent builds, another reviews, an inspector checks the result. Anything that can't be undone waits for a person.

Proof The lab notebook

Tested, then trusted.

Nothing goes into production on a hunch. We test it on real work and write down what happened, including when the answer is no.

1. September 2026

   Can a small decision model out-file a big local one?

   Yes. 77% exact matches against 66 to 68% on 84 hand-checked items. At a 0.75 confidence gate it files half of everything at 95% precision and asks for help on the rest.

2. September 2026

   Can a rebuild be simpler than the original?

   Yes. The intake system was rewritten after its predecessor collapsed under a sixteen-section rulebook. The new one shipped in a day and is still running.

3. October 2026

   Can a tiny model learn the news by training on it daily?

   Not yet. It learns, and it forgets like the news cycle, but even a 12-billion-parameter model couldn't answer questions about what it read. For now, looking it up beats remembering it.

Work with us

Bring us one workflow.

The one that keeps eating your week. Show us what comes in, what has to come out and who has to trust the result, and we'll find the first system worth building.

1. Map itWe follow one piece of work from what comes in to what goes out: where the time goes, what should be plain, dependable software, where judgment is needed and what must stay with a person. You get the map, the design and a plan for a pilot.
2. Build one laneWe build that workflow end to end: dependable software first, agents where judgment is needed, tests and monitoring from day one. Then we measure it against the way you worked before.
3. Run it with youOn your server or in your cloud, with keys locked away and a person deciding anything that can't be undone. When it earns its place, we grow it.