How a Bitcoin infrastructure company runs marketing, CRM, and operations on ~60 governed AI agents — and the hub page for every deep-dive we've published about the system.

Most marketing teams using AI are stuck in one of two places: re-explaining their company to a chatbot every morning, or juggling a stack of AI tools that don't talk to each other. We were too. This article explains the system we built instead: one workspace that already knows the business, with around 60 governed agents running on top of it. It also maps every deep-dive we've published about how it works.

BitSafe is an infrastructure company. We brought Bitcoin onto Canton with CBTC, and we're open-sourcing the Decentralization Manager. In the simplest framing, we make it easier to build decentralized tech on Canton. At some point we realized the instinct that drives our products could run our own team too: build the layer others build on. That's where this system came from.

The three levels of AI

Level 1: an LLM in a browser tab. ChatGPT, Claude, Gemini. Powerful, but it starts every task with amnesia. You re-upload the context folder and re-explain the company, every single time.

Level 2: a stack of AI point tools. Faster, but your context is scattered across tools that don't talk to each other. None of them knows your business end to end. A good place to pass through; a bad place to stop.

Level 3: one system that already holds your context. The AI starts every task knowing who you are, what you do, and what you're selling. The discipline behind this is context engineering: instead of perfecting prompts, you manage what the AI already knows.

The rest of this page is how we climbed to Level 3.

flowchart TB
    L1["Level 1 — An LLM in a browser tab<br>Powerful, but starts every task with amnesia"]
    L2["Level 2 — A stack of AI point tools<br>Faster, but context is scattered across tools that don't talk"]
    L3["Level 3 — One system that already holds your context<br>AI starts every task knowing the business"]
    L1 -->|"add tools"| L2
    L2 -->|"add context engineering"| L3
    style L3 fill:#F4652F,color:#fff

The system: a brain and a reach

Notion is the brain, and where the work gets made. Brand guidelines, CRM, SOPs, meeting notes, rules: everything the company knows lives in one structured workspace. Our agents run on top of it with that context preloaded, so output sounds like us without a paragraph of setup.

NanoClaw is the reach. When a task needs something from outside the workspace, NanoClaw goes and gets it from Slack, the web, transcripts, or on-chain data, then hands what it finds back to our Notion agents, which produce the work. Retrieve, hand off, produce.

flowchart LR
    Slack["Slack"] --> NC
    Web["Web"] --> NC
    Calls["Call transcripts"] --> NC
    Chain["On-chain data"] --> NC
    subgraph NC["NanoClaw — the reach"]
        Jobs["~80 scheduled jobs<br>retrieve and pre-process"]
    end
    subgraph Brain["Notion — the brain"]
        Context["CRM · Docs · SOPs · Meeting notes"]
        Agents["~60 governed agents"]
        Context --- Agents
    end
    NC -->|"hand off"| Brain
    Brain -->|"human approves"| Out["Work ships"]
    Brain -.->|"each retrieval compounds the context"| Context
    style Brain fill:#F4652F,color:#fff

They stack — they don't compete. The internal rule of thumb fits on a sticky note: try the brain first; reach for the hands when the brain can't. Finding or summarizing internal knowledge, running a routine workflow, answering a question grounded in company content — that's the workspace, because it's simpler, auditable, and already holds the context. Pulling live data from outside the walls, solving an open-ended problem that spans many sources, spinning up a one-off tool — that's the reach. The division of labor is clean enough that nobody on the team has to think about it.

And the seam is invisible. The two systems are wired together, so a workspace agent that hits the edge of what it knows can ask the reach for what it needs mid-task. The person who made the request never has to know which system did the work. The best workflows use both on purpose: the reach fetches and pre-processes external data, and the brain structures it, stores it, and folds it into what every future task already knows. Each retrieval makes the system permanently smarter. That's the compounding loop a stack of disconnected point tools can never give you.

A worked example: our weekly content system. It listens to client calls, reads our Slack channels, and pulls from the web to recommend content each week. Once we approve a topic, it schedules it, writes the draft, and prepares a design request. After design, it reviews the final asset, and a human gives the last sign-off before anything ships.

Operations first, output second

The counterintuitive lesson: most of what our ~60 governed agents do isn't writing content. It's operations: CRM hygiene and enrichment, meeting prep and briefs, monitoring, reporting, document verification.

Start with AI on operations and the way work gets done improves immediately. Build those workflows one by one, and the system accumulates enough context that good output becomes the bonus. It sounds like you, because the context lives in the system rather than in anyone's prompt.

What ~60 governed agents actually buy you

"Sixty agents" can sound abstract, so here's the practical effect. Almost every agent does one small job: classify the new CRM row, prep tomorrow's meetings, flag the document whose verification expired, nudge the owner of a stalled draft, score the pipeline before Monday morning. Individually, none of it is impressive. Together, three things change:

• The routine work stops slipping. The enrichment, hygiene checks, briefs, and reminders everyone agrees matter, and that nobody was ever going to do by hand, just happen now. One contact-classification agent alone has run about 7,700 times.
• Quality stops depending on who's busy. Every company gets classified against the same brand-guide definitions. Every meeting gets prepped to the same standard. Every document gets re-verified on the same cycle. Consistency is the quiet benefit nobody puts on a slide.
• People keep the judgment calls. Anything outbound stops at a human approval gate: agents propose, people approve. And no agent can edit its own instructions: every change is proposed, approved by a human owner, and applied through a controlled process with a full audit trail.

How we keep ~60 of them governable (the registry, the change control, the agents that govern the agents) is a story of its own, and it gets a full deep-dive in the reading map below.

The path to Level 3

1. Get the knowledge in one place. Our migration started in November 2025; we moved documents over as we used them rather than all at once. Data hygiene matters as much as volume. We verify documents on a review cycle and archive what no longer earns its place, so the AI isn't writing from junk.
2. Point the AI at one painful workflow. Don't boil the ocean. Pick the workflow everyone dreads and make the system genuinely good at it.
3. Build governed agents one by one, with a human in the loop. Every agent has an owner, instructions under change control, and approval gates where it matters.

The unlock was never the model. It was the context.

The reading map

Everything below goes a level deeper on one part of the system.

How BitSafe runs on Notion — the brain:

The NanoClaw series — the reach:

Standalone deep-dives — the seams and the scoreboard:

The source repo. The article series — drafts, worked examples, and the case-study docs behind it — lives in the open at github.com/Akibalogh/bitsafe-ai-docs, built on the upstream NanoClaw framework.

Where to start

Stop treating your workspace like a filing cabinet. Treat it like a teammate that knows the business, and give it the best context you can.

If you're building on Canton and want to compare notes on marketing operations, find Kadeem Clarke on the Canton ecosystem Slack. Happy to walk through any part of the setup.