Connectivity was the last mile. Then it needed an operator.

Anthropic put it plainly when they acquired Stainless: the frontier of AI is shifting from models that answer to agents that act — and agents are only as capable as the systems they can reach. They priced that reach at over $300 million. Reaching your systems is real, hard, valuable work.

So the industry has built a lot to close that last mile: REST APIs, API gateways, generated SDKs, MCP servers. All of them do the same essential job — they make your systems reachable by an AI. We built in this category too. dataBridges handles real-time communication between distributed systems; APIFront is an API gateway that puts an organisation's internal functions safely in reach of modern AI, with the security, authentication, and monitoring an enterprise actually needs. The last mile of connectivity — closed.

But once an agent can reach a system, the next question is whether it can operate that system correctly. That is a different problem entirely.

Tool access tells an AI what it can do. It says nothing about what it should do — in what order, under what rules, for whom. A dispatched order can't be cancelled. KYC must clear before activation. This role can read but not write. Those rules have to live somewhere. If they don't live in the system, they live nowhere — and an AI with full tool access and no operational knowledge is just a faster way to break things.

So a reachable system needs one more thing: a reasoning layer trained to operate it correctly. Not a wider bridge — an operator standing on the bridge who knows the rules.

Building that operator is its own problem. The default approach loads a single AI with two jobs at once: hold the conversation, and figure out how to drive the tools correctly. Fine for a handful of simple tools. It breaks the moment the system is real — sequences, state, access boundaries, exceptions, data. The reasoning that handles a customer conversation is not the reasoning that decides whether KYC precedes activation. Load one layer with both and you get behaviour you can't audit.

And the data is its own boundary. A reasoning layer doesn't need every row to answer a question — it needs the answer. The operator pulls what's relevant, filters it against who's asking and what they're allowed to see, and returns only the conclusion. The AI reasons over conclusions, not raw data. Security, audit, and cost — handled in one architectural choice.

So a real toolsystem needs its own agent: one that holds the rules of correct use, reasons at the depth those rules demand, and exposes a conversational surface other AIs can talk to.

We built that too. Nexus is a toolsystem agent (powered by Claude) — the operator that stands on the connectivity and drives the tools correctly. The upstream AI handles the conversation; Nexus handles correct use, and decides what data crosses back to the model. It sits on APIFront and dataBridges — the last mile we'd already closed, now with an operator who knows how to use it.

Reaching a system and operating it correctly are two different problems. Connectivity solved the first. Everything else turns on the second — not whether your AI can reach your systems, but whether it operates them correctly.

That's the realization the rest of these essays build on.