Codified scheduling patterns can turn tools into hidden schedulers

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

As agent applications mature, some repeated decisions stop needing open-ended judgment and codify into deterministic procedure.

Many codified patterns fit naturally inside ordinary tools — formatting, parsing, validation, data transformation. These patterns describe how to perform a single capability, and they improve when they move from prompt instructions into exact code.

A deterministic multi-step implementation can still be an ordinary tool when those steps remain internal to one coherent capability rather than choosing what broader task step should happen next.

Scheduling patterns are different. They describe what should happen next:

  • after editing, always run tests before summarizing
  • decompose feature work into research, implementation, and review
  • after a failed tool path, retry once with narrower context and then escalate
  • after collecting search results, deduplicate and re-rank before presenting

When a tool embeds next-step policy across subtasks, retries, recursion, or tool-surface selection, it stops acting like an ordinary capability and starts acting like a hidden scheduler. A run_feature_workflow tool that decides which subtask to run, which tool surface to expose, when to recurse, and when to stop is not a tool — it is an orchestration runtime behind a tool-shaped facade.

This works, but it is not where codified scheduling wants to live. Once next-step policy has stabilized enough to become code, the application needs to replace model-chosen transitions with explicit control logic. When codified policy needs to choose the next subtask, alter the next tool surface, or manage retries and stopping conditions across task boundaries, a framework that exposes tools but not progression forces that control logic to masquerade as tool implementation.


Relevant Notes: