An engineering note on the agent loop and the learning loop

This is not a product launch. It describes two Python packages in the OmniBot tree: agent_engine and learning_engine. Both run unit tests without a robot. Neither updates a fleet, and neither uses an NPU.

agent_engine

AgentHarness advances one state per tick(): idle, perceive, plan, act, monitor, reflect, remember. A step budget and an emergency-stop interrupt are part of the loop. Low-level tools are supposed to pass a verifier before they run.

The verifier port wraps hardware-limit checks (speed and reach) that live next to the learning engine. Tests drive the harness with a scripted reasoner, so the loop can be exercised without a cloud model.

learning_engine

The learning package is a post-training loop: collect episodes, score them, replay them, and optionally hand training off to another engine. The numpy core does not need a GPU. A safety check rejects actions outside the robot's published speed and joint limits.

Hybrid cloud / on-device / NPU routing is roadmap. Over-the-air delivery to robots already deployed is roadmap. What you can run today is the test suite.

On a checkout of the public OmniBot repository, the agent tests live under omnibot-ai-engines/agent_engine and the learning tests under omnibot-ai-engines/learning_engine.