From Prompts to Physical Action

The moment an AI system can act, the design problem changes. Capability is no longer enough. We need to know what the agent is allowed to do, how uncertainty is handled and when a human must intervene.

Three layers of action

1. Reasoning: understand the goal, context and constraints.
2. Execution: call a tool, change a system or move something in the world.
3. Verification: independently confirm what actually changed.

Google DeepMind describes Gemini Robotics 2 as an intelligence layer for whole-body control, dexterity and multi-robot collaboration. Its safety work includes uncertainty handling, unsafe tool-call refusal and requests for human intervention. Google DeepMind's official overview.

OpenAI presents Codex as a command centre for long-running agent work across building, testing and maintaining software. The human operator still defines the task, reviews changes and controls the environment in which the agent works. OpenAI's Codex app announcement.

The same lesson applies to commerce

An agent preparing a merchandising brief is different from one changing prices. Drafting a campaign is different from enrolling customers. Finding a supplier is different from offering a product for sale. Each boundary needs its own authority and receipt.

Agentic does not mean autonomous by default. It means capability paired with explicit authority.