AI as the Control Plane: Orchestrating Robots, Satellites and Payments

The agent is the easy part

Our evidence report on AI, robotics, space and Bitcoin describes four technologies as layers of one system, with AI as the control plane. The figures behind it are public: 78% of organisations reported using AI in 2024, and 5 million industrial robots were in operation at the end of 2025. What those figures do not show is the engineering between the layers.

A language model in a loop can plan a delivery in seconds. Making that plan safe to execute, when a tool call can start a robot or release a payment, is where the work is. Four design decisions carry most of it.

1. Tools with contracts

An agent acts only through tools, so each tool is a contract: typed inputs, typed outputs, and a clear statement of what it changes in the world.

• Separate read tools from write tools. Reading fleet status is free; dispatching a robot is not.
• Make write tools idempotent, so a retried call cannot dispatch the same pick twice.
• Return structured errors the model can reason about, not free text.

2. State that survives a dropped link

Physical work outlasts a model’s context window and any single network connection. Satellite links extend coverage, but they do not remove delay or outages.

• Keep the job’s state in a durable store, not in the conversation.
• Treat telemetry as events with timestamps, and decide explicitly what to do when they arrive late or out of order.
• Define a safe default for silence: a robot that hears nothing stops.

3. Limits on money and motion

Protocols such as L402 make it technically possible for a program to pay for a service. That is a reason for limits, not for trust.

• Give each agent a spending cap per job and per day, enforced outside the model.
• Require a person to approve any step above a threshold, and record who approved it.
• Log every tool call with its inputs, outputs and cost, so the run can be audited afterwards.

4. Evaluation before autonomy

An agent earns autonomy the way a model earns production: by passing tests that someone wrote down in advance.

• Replay recorded jobs and compare the agent’s plan with what a competent operator did.
• Inject failures on purpose: a blocked aisle, a lost link, a rejected payment.
• Track the same metrics in production, and roll back when they move.

Where to start

Start with one workflow and read-only tools. Let the agent propose plans that a person executes, and measure how often the proposals are accepted unchanged. Add write tools one at a time, each behind an approval step, once that number is high enough to justify it.

The full analysis, with the sources for every figure and five step-through diagrams of the underlying algorithms, is in the evidence report.