All research notes

Loop Engineering

The Cost-Aware Agent Loop

Jun 28, 2026 · 5 min read

Most agent stacks pick a model once, at design time, and then pay that model's price for every step regardless of difficulty. Classification, extraction, and formatting steps do not need frontier reasoning; planning and repair often do.

Treating the loop as a controller lets us attach a budget to a task rather than a model to a workflow. Each step emits a difficulty estimate, the router selects the cheapest model whose historical success rate clears the threshold, and escalation happens only on verifier failure.

In enterprise workloads we consistently see 60–80% of steps safely served by small models, with end-to-end quality held flat. The savings come from the distribution of easy steps, not from downgrading the hard ones.

The operational requirement is per-step telemetry: success rate, latency, and cost per model per step type. Without it, routing is guesswork dressed as optimization.