Agent loops, tools and skills — defined on the Anthropic stack, and named for the CFO's work.
Falling token prices do not make agentic AI cheaper, because a loop buys many units of model per task where a prompt buys one. That single fact moves agent design out of architecture and onto the CFO's desk: what a loop may reach, what it may do without asking, and what it costs to run are now finance questions. Across four essays we take the loop apart in plain language — then name the CFO's version of it.
Tuesday, August 18 · about 14 min
Token prices are falling. Agent inference costs are rising. The difference is who runs the loop.
What an agent may reach, what it may change, and what it must ask permission for.
Packaged procedure, loaded on demand — and the reason your context window is a budget line.
The finance function's own named library — one loop per unit of work, one skill per procedure.
Where the loops land first in a carrier's finance function, and what a day of DSO is worth.
Consumption pricing, deferred revenue and the finance function that meters its own product.
Parts 01–04 are live; 2 more follow on the dates shown. Each essay links to a full-length analysis — the technical layer behind the argument.
The framing of this series — a harness taken apart into six named parts, a single run traced from start to finish, and a catalog of named skills with an orchestrating layer above them — is inspired by dadloop, an agent harness built and published by Swami Chandrasekaran, Partner and Global Head of AI & Data Labs at KPMG (LinkedIn · github.com/swamichandra/dadloop). He explained a harness through Dad. We are explaining one through the close. The teaching devices are his; the finance domain, the loop library and every figure here are ours.