EELML STUDIO PERSONAL 0.18
How the model-owned Agent Loop works
In EELML, the active model owns semantic decisions. Retrieval ranks several relevant capabilities, but code does not silently choose the final tool. The model selects a capability and arguments; the Agent Kernel validates and executes; the result or failure returns as a typed Observation for the model to evaluate.
Updated 2026-10-02 · This page describes the current Personal 0.18 release.
One bounded work cycle
- The user gives a goal in a chat or workspace-bound work conversation.
- Tool Search retrieves several relevant capability candidates.
- The active model chooses whether to answer directly or call a tool.
- The Agent Kernel validates permissions and arguments, then executes.
- A typed Observation returns to the same model.
- The model can finish, repair arguments, choose another capability, or ask the user.
Why failure returns to the model
A tool error is evidence, not a hidden reason to switch paths. EELML normalizes failures into typed results so the model can explain the issue or make a bounded recovery decision.
What the Agent does not own
The Agent Kernel does not load model weights and does not contain UI implementation. The native App does not duplicate tool routing, DAG logic, or model policy. These boundaries keep execution inspectable and independently testable.