Stateful tools recover control by becoming hidden schedulers

Type: note · Status: seedling · Tags: computational-model, context-engineering, tool-loop

A naive version of the "expose the loop" argument claims that a framework-owned tool loop cannot express what application-owned orchestration requires. That version is false. If we allow a sufficiently stateful tool — a singleton runtime behind the tool boundary — a hidden loop can recover substantial control.

The recovery is genuine. Such a runtime can hold retries, checkpoints, branch records, and recursion state — enough orchestration logic that the model never has to invoke a visible scheduler. From the model's perspective nothing has changed; from the application's perspective, the tool is quietly running the show.

But notice what happened. The scheduler — the code deciding what happens next — has been relocated into a special tool. The framework loop is still running, but it is no longer the locus of control. A tool became a covert orchestration runtime.

That reframes the question. The hard problem is not whether hidden loops can recover expressivity — they can. It is where the scheduler lives and whether the framework exposes it or forces it to masquerade as tool implementation. And even this recovery has limits: changing the action alphabet between sub-tasks and sub-goals that exceed one context window are the strongest cases where a hidden scheduler starts to buckle.


Relevant Notes: