Bigger LLM Coordinator Fails to Fix Agent Loop; Deterministic Script Does

A series installment reports that upgrading to a larger coordinator model did not resolve orchestration failures such as dropped steps, delegations emitted as text, and empty turns after tool calls. The author argues the root cause is the agentic loop itself, since the model must simultaneously plan, dispatch, track outstanding work, and evaluate results. Handing the control loop to a small deterministic script, with the model used only for planning and evaluation, produced immediate gains. In the reported test, a plan that had stalled after three of five steps completed fully once executed by the script.