Agent platforms
Long tool-using loops, the same task shapes over and over, and a bill dominated by context you are re-sending.
An agent platform is the workload Operant was built against, because it has all three cost pathologies at once.
Context rot in tool loops
Each tool result is appended and then re-sent on every subsequent turn. By step twelve, most of the prompt is the transcript of steps one through eleven, and a large share of it no longer affects the answer. Compression here is KV-compatible, so the prompt shrinks without discarding the cache that makes it cheap.
Uniform model choice across non-uniform work
A single loop mixes planning, tool selection, and summarising what came back. Those are not the same difficulty, and paying the planning tier for all three is the most common overspend we see. The Gateway assigns per call — and accounts for the fact that switching models breaks the cache.
Repetition you are paying full price for
Platform workloads repeat heavily: the same handful of task shapes, thousands of times a day, each one re-deriving what the last one already established. Cortex learns those into skills.
Where to start
A waste audit answers this with your numbers rather than ours. Nothing to install, and you keep the findings either way.
This page describes where cost concentrates in this kind of workload. It is not a claim about customers in this sector — Operant is pre-launch and has none to name.